Adding upstream version 86.0.1.upstream/86.0.1upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

57 files changed, 9767 insertions, 0 deletions

diff --git a/memory/replace/dmd/DMD.cpp b/memory/replace/dmd/DMD.cpp
new file mode 100644
index 0000000000..f93c8eebca
--- /dev/null
+++ b/memory/replace/dmd/DMD.cpp
@@ -0,0 +1,1882 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <ctype.h>
+#include <errno.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#if !defined(MOZ_PROFILING)
+#  error "DMD requires MOZ_PROFILING"
+#endif
+
+#ifdef XP_WIN
+#  include <windows.h>
+#  include <process.h>
+#else
+#  include <pthread.h>
+#  include <sys/types.h>
+#  include <unistd.h>
+#endif
+
+#ifdef ANDROID
+#  include <android/log.h>
+#endif
+
+#include "nscore.h"
+
+#include "mozilla/Assertions.h"
+#include "mozilla/FastBernoulliTrial.h"
+#include "mozilla/HashFunctions.h"
+#include "mozilla/HashTable.h"
+#include "mozilla/IntegerPrintfMacros.h"
+#include "mozilla/JSONWriter.h"
+#include "mozilla/Likely.h"
+#include "mozilla/MemoryReporting.h"
+#include "mozilla/PodOperations.h"
+#include "mozilla/StackWalk.h"
+#include "mozilla/ThreadLocal.h"
+
+// CodeAddressService is defined entirely in the header, so this does not make
+// DMD depend on XPCOM's object file.
+#include "CodeAddressService.h"
+
+// replace_malloc.h needs to be included before replace_malloc_bridge.h,
+// which DMD.h includes, so DMD.h needs to be included after replace_malloc.h.
+#include "replace_malloc.h"
+#include "DMD.h"
+
+namespace mozilla {
+namespace dmd {
+
+class DMDBridge : public ReplaceMallocBridge {
+  virtual DMDFuncs* GetDMDFuncs() override;
+};
+
+static DMDBridge* gDMDBridge;
+static DMDFuncs gDMDFuncs;
+
+DMDFuncs* DMDBridge::GetDMDFuncs() { return &gDMDFuncs; }
+
+MOZ_FORMAT_PRINTF(1, 2)
+inline void StatusMsg(const char* aFmt, ...) {
+  va_list ap;
+  va_start(ap, aFmt);
+  gDMDFuncs.StatusMsg(aFmt, ap);
+  va_end(ap);
+}
+
+//---------------------------------------------------------------------------
+// Utilities
+//---------------------------------------------------------------------------
+
+#ifndef DISALLOW_COPY_AND_ASSIGN
+#  define DISALLOW_COPY_AND_ASSIGN(T) \
+    T(const T&);                      \
+    void operator=(const T&)
+#endif
+
+static malloc_table_t gMallocTable;
+
+// This provides infallible allocations (they abort on OOM).  We use it for all
+// of DMD's own allocations, which fall into the following three cases.
+//
+// - Direct allocations (the easy case).
+//
+// - Indirect allocations in mozilla::{Vector,HashSet,HashMap} -- this class
+//   serves as their AllocPolicy.
+//
+// - Other indirect allocations (e.g. MozStackWalk) -- see the comments on
+//   Thread::mBlockIntercepts and in replace_malloc for how these work.
+//
+// It would be nice if we could use the InfallibleAllocPolicy from mozalloc,
+// but DMD cannot use mozalloc.
+//
+class InfallibleAllocPolicy {
+  static void ExitOnFailure(const void* aP);
+
+ public:
+  template <typename T>
+  static T* maybe_pod_malloc(size_t aNumElems) {
+    if (aNumElems & mozilla::tl::MulOverflowMask<sizeof(T)>::value)
+      return nullptr;
+    return (T*)gMallocTable.malloc(aNumElems * sizeof(T));
+  }
+
+  template <typename T>
+  static T* maybe_pod_calloc(size_t aNumElems) {
+    return (T*)gMallocTable.calloc(aNumElems, sizeof(T));
+  }
+
+  template <typename T>
+  static T* maybe_pod_realloc(T* aPtr, size_t aOldSize, size_t aNewSize) {
+    if (aNewSize & mozilla::tl::MulOverflowMask<sizeof(T)>::value)
+      return nullptr;
+    return (T*)gMallocTable.realloc(aPtr, aNewSize * sizeof(T));
+  }
+
+  static void* malloc_(size_t aSize) {
+    void* p = gMallocTable.malloc(aSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static T* pod_malloc(size_t aNumElems) {
+    T* p = maybe_pod_malloc<T>(aNumElems);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  static void* calloc_(size_t aCount, size_t aSize) {
+    void* p = gMallocTable.calloc(aCount, aSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static T* pod_calloc(size_t aNumElems) {
+    T* p = maybe_pod_calloc<T>(aNumElems);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  static void* realloc_(void* aPtr, size_t aNewSize) {
+    void* p = gMallocTable.realloc(aPtr, aNewSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static T* pod_realloc(T* aPtr, size_t aOldSize, size_t aNewSize) {
+    T* p = maybe_pod_realloc(aPtr, aOldSize, aNewSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  static void* memalign_(size_t aAlignment, size_t aSize) {
+    void* p = gMallocTable.memalign(aAlignment, aSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static void free_(T* aPtr, size_t aSize = 0) {
+    gMallocTable.free(aPtr);
+  }
+
+  static char* strdup_(const char* aStr) {
+    char* s = (char*)InfallibleAllocPolicy::malloc_(strlen(aStr) + 1);
+    strcpy(s, aStr);
+    return s;
+  }
+
+  template <class T>
+  static T* new_() {
+    void* mem = malloc_(sizeof(T));
+    return new (mem) T;
+  }
+
+  template <class T, typename P1>
+  static T* new_(const P1& aP1) {
+    void* mem = malloc_(sizeof(T));
+    return new (mem) T(aP1);
+  }
+
+  template <class T>
+  static void delete_(T* aPtr) {
+    if (aPtr) {
+      aPtr->~T();
+      InfallibleAllocPolicy::free_(aPtr);
+    }
+  }
+
+  static void reportAllocOverflow() { ExitOnFailure(nullptr); }
+  bool checkSimulatedOOM() const { return true; }
+};
+
+// This is only needed because of the |const void*| vs |void*| arg mismatch.
+static size_t MallocSizeOf(const void* aPtr) {
+  return gMallocTable.malloc_usable_size(const_cast<void*>(aPtr));
+}
+
+void DMDFuncs::StatusMsg(const char* aFmt, va_list aAp) {
+#ifdef ANDROID
+  __android_log_vprint(ANDROID_LOG_INFO, "DMD", aFmt, aAp);
+#else
+  // The +64 is easily enough for the "DMD[<pid>] " prefix and the NUL.
+  char* fmt = (char*)InfallibleAllocPolicy::malloc_(strlen(aFmt) + 64);
+  sprintf(fmt, "DMD[%d] %s", getpid(), aFmt);
+  vfprintf(stderr, fmt, aAp);
+  InfallibleAllocPolicy::free_(fmt);
+#endif
+}
+
+/* static */
+void InfallibleAllocPolicy::ExitOnFailure(const void* aP) {
+  if (!aP) {
+    MOZ_CRASH("DMD out of memory; aborting");
+  }
+}
+
+static double Percent(size_t part, size_t whole) {
+  return (whole == 0) ? 0 : 100 * (double)part / whole;
+}
+
+// Commifies the number.
+static char* Show(size_t n, char* buf, size_t buflen) {
+  int nc = 0, i = 0, lasti = buflen - 2;
+  buf[lasti + 1] = '\0';
+  if (n == 0) {
+    buf[lasti - i] = '0';
+    i++;
+  } else {
+    while (n > 0) {
+      if (((i - nc) % 3) == 0 && i != 0) {
+        buf[lasti - i] = ',';
+        i++;
+        nc++;
+      }
+      buf[lasti - i] = static_cast<char>((n % 10) + '0');
+      i++;
+      n /= 10;
+    }
+  }
+  int firstCharIndex = lasti - i + 1;
+
+  MOZ_ASSERT(firstCharIndex >= 0);
+  return &buf[firstCharIndex];
+}
+
+//---------------------------------------------------------------------------
+// Options (Part 1)
+//---------------------------------------------------------------------------
+
+class Options {
+  template <typename T>
+  struct NumOption {
+    const T mDefault;
+    const T mMax;
+    T mActual;
+    NumOption(T aDefault, T aMax)
+        : mDefault(aDefault), mMax(aMax), mActual(aDefault) {}
+  };
+
+  // DMD has several modes. These modes affect what data is recorded and
+  // written to the output file, and the written data affects the
+  // post-processing that dmd.py can do.
+  //
+  // Users specify the mode as soon as DMD starts. This leads to minimal memory
+  // usage and log file size. It has the disadvantage that is inflexible -- if
+  // you want to change modes you have to re-run DMD. But in practice changing
+  // modes seems to be rare, so it's not much of a problem.
+  //
+  // An alternative possibility would be to always record and output *all* the
+  // information needed for all modes. This would let you choose the mode when
+  // running dmd.py, and so you could do multiple kinds of profiling on a
+  // single DMD run. But if you are only interested in one of the simpler
+  // modes, you'd pay the price of (a) increased memory usage and (b) *very*
+  // large log files.
+  //
+  // Finally, another alternative possibility would be to do mode selection
+  // partly at DMD startup or recording, and then partly in dmd.py. This would
+  // give some extra flexibility at moderate memory and file size cost. But
+  // certain mode pairs wouldn't work, which would be confusing.
+  //
+  enum class Mode {
+    // For each live block, this mode outputs: size (usable and slop) and
+    // (possibly) and allocation stack. This mode is good for live heap
+    // profiling.
+    Live,
+
+    // Like "Live", but for each live block it also outputs: zero or more
+    // report stacks. This mode is good for identifying where memory reporters
+    // should be added. This is the default mode.
+    DarkMatter,
+
+    // Like "Live", but also outputs the same data for dead blocks. This mode
+    // does cumulative heap profiling, which is good for identifying where large
+    // amounts of short-lived allocations ("heap churn") occur.
+    Cumulative,
+
+    // Like "Live", but this mode also outputs for each live block the address
+    // of the block and the values contained in the blocks. This mode is useful
+    // for investigating leaks, by helping to figure out which blocks refer to
+    // other blocks. This mode force-enables full stacks coverage.
+    Scan
+  };
+
+  // With full stacks, every heap block gets a stack trace recorded for it.
+  // This is complete but slow.
+  //
+  // With partial stacks, not all heap blocks will get a stack trace recorded.
+  // A Bernoulli trial (see mfbt/FastBernoulliTrial.h for details) is performed
+  // for each heap block to decide if it gets one. Because bigger heap blocks
+  // are more likely to get a stack trace, even though most heap *blocks* won't
+  // get a stack trace, most heap *bytes* will.
+  enum class Stacks { Full, Partial };
+
+  char* mDMDEnvVar;  // a saved copy, for later printing
+
+  Mode mMode;
+  Stacks mStacks;
+  bool mShowDumpStats;
+
+  void BadArg(const char* aArg);
+  static const char* ValueIfMatch(const char* aArg, const char* aOptionName);
+  static bool GetLong(const char* aArg, const char* aOptionName, long aMin,
+                      long aMax, long* aValue);
+  static bool GetBool(const char* aArg, const char* aOptionName, bool* aValue);
+
+ public:
+  explicit Options(const char* aDMDEnvVar);
+
+  bool IsLiveMode() const { return mMode == Mode::Live; }
+  bool IsDarkMatterMode() const { return mMode == Mode::DarkMatter; }
+  bool IsCumulativeMode() const { return mMode == Mode::Cumulative; }
+  bool IsScanMode() const { return mMode == Mode::Scan; }
+
+  const char* ModeString() const;
+
+  const char* DMDEnvVar() const { return mDMDEnvVar; }
+
+  bool DoFullStacks() const { return mStacks == Stacks::Full; }
+  size_t ShowDumpStats() const { return mShowDumpStats; }
+};
+
+static Options* gOptions;
+
+//---------------------------------------------------------------------------
+// The global lock
+//---------------------------------------------------------------------------
+
+// MutexBase implements the platform-specific parts of a mutex.
+
+#ifdef XP_WIN
+
+class MutexBase {
+  CRITICAL_SECTION mCS;
+
+  DISALLOW_COPY_AND_ASSIGN(MutexBase);
+
+ public:
+  MutexBase() { InitializeCriticalSection(&mCS); }
+  ~MutexBase() { DeleteCriticalSection(&mCS); }
+
+  void Lock() { EnterCriticalSection(&mCS); }
+  void Unlock() { LeaveCriticalSection(&mCS); }
+};
+
+#else
+
+class MutexBase {
+  pthread_mutex_t mMutex;
+
+  MutexBase(const MutexBase&) = delete;
+
+  const MutexBase& operator=(const MutexBase&) = delete;
+
+ public:
+  MutexBase() { pthread_mutex_init(&mMutex, nullptr); }
+
+  void Lock() { pthread_mutex_lock(&mMutex); }
+  void Unlock() { pthread_mutex_unlock(&mMutex); }
+};
+
+#endif
+
+class Mutex : private MutexBase {
+  bool mIsLocked;
+
+  Mutex(const Mutex&) = delete;
+
+  const Mutex& operator=(const Mutex&) = delete;
+
+ public:
+  Mutex() : mIsLocked(false) {}
+
+  void Lock() {
+    MutexBase::Lock();
+    MOZ_ASSERT(!mIsLocked);
+    mIsLocked = true;
+  }
+
+  void Unlock() {
+    MOZ_ASSERT(mIsLocked);
+    mIsLocked = false;
+    MutexBase::Unlock();
+  }
+
+  bool IsLocked() { return mIsLocked; }
+};
+
+// This lock must be held while manipulating global state such as
+// gStackTraceTable, gLiveBlockTable, gDeadBlockTable. Note that gOptions is
+// *not* protected by this lock because it is only written to by Options(),
+// which is only invoked at start-up and in ResetEverything(), which is only
+// used by SmokeDMD.cpp.
+static Mutex* gStateLock = nullptr;
+
+class AutoLockState {
+  AutoLockState(const AutoLockState&) = delete;
+
+  const AutoLockState& operator=(const AutoLockState&) = delete;
+
+ public:
+  AutoLockState() { gStateLock->Lock(); }
+  ~AutoLockState() { gStateLock->Unlock(); }
+};
+
+class AutoUnlockState {
+  AutoUnlockState(const AutoUnlockState&) = delete;
+
+  const AutoUnlockState& operator=(const AutoUnlockState&) = delete;
+
+ public:
+  AutoUnlockState() { gStateLock->Unlock(); }
+  ~AutoUnlockState() { gStateLock->Lock(); }
+};
+
+//---------------------------------------------------------------------------
+// Per-thread blocking of intercepts
+//---------------------------------------------------------------------------
+
+// On MacOS, the first __thread/thread_local access calls malloc, which leads
+// to an infinite loop. So we use pthread-based TLS instead, which somehow
+// doesn't have this problem.
+#if !defined(XP_DARWIN)
+#  define DMD_THREAD_LOCAL(T) MOZ_THREAD_LOCAL(T)
+#else
+#  define DMD_THREAD_LOCAL(T) \
+    detail::ThreadLocal<T, detail::ThreadLocalKeyStorage>
+#endif
+
+class Thread {
+  // Required for allocation via InfallibleAllocPolicy::new_.
+  friend class InfallibleAllocPolicy;
+
+  // When true, this blocks intercepts, which allows malloc interception
+  // functions to themselves call malloc.  (Nb: for direct calls to malloc we
+  // can just use InfallibleAllocPolicy::{malloc_,new_}, but we sometimes
+  // indirectly call vanilla malloc via functions like MozStackWalk.)
+  bool mBlockIntercepts;
+
+  Thread() : mBlockIntercepts(false) {}
+
+  Thread(const Thread&) = delete;
+
+  const Thread& operator=(const Thread&) = delete;
+
+  static DMD_THREAD_LOCAL(Thread*) tlsThread;
+
+ public:
+  static void Init() {
+    if (!tlsThread.init()) {
+      MOZ_CRASH();
+    }
+  }
+
+  static Thread* Fetch() {
+    Thread* t = tlsThread.get();
+    if (MOZ_UNLIKELY(!t)) {
+      // This memory is never freed, even if the thread dies. It's a leak, but
+      // only a tiny one.
+      t = InfallibleAllocPolicy::new_<Thread>();
+      tlsThread.set(t);
+    }
+
+    return t;
+  }
+
+  bool BlockIntercepts() {
+    MOZ_ASSERT(!mBlockIntercepts);
+    return mBlockIntercepts = true;
+  }
+
+  bool UnblockIntercepts() {
+    MOZ_ASSERT(mBlockIntercepts);
+    return mBlockIntercepts = false;
+  }
+
+  bool InterceptsAreBlocked() const { return mBlockIntercepts; }
+};
+
+DMD_THREAD_LOCAL(Thread*) Thread::tlsThread;
+
+// An object of this class must be created (on the stack) before running any
+// code that might allocate.
+class AutoBlockIntercepts {
+  Thread* const mT;
+
+  AutoBlockIntercepts(const AutoBlockIntercepts&) = delete;
+
+  const AutoBlockIntercepts& operator=(const AutoBlockIntercepts&) = delete;
+
+ public:
+  explicit AutoBlockIntercepts(Thread* aT) : mT(aT) { mT->BlockIntercepts(); }
+  ~AutoBlockIntercepts() {
+    MOZ_ASSERT(mT->InterceptsAreBlocked());
+    mT->UnblockIntercepts();
+  }
+};
+
+//---------------------------------------------------------------------------
+// Location service
+//---------------------------------------------------------------------------
+
+struct DescribeCodeAddressLock {
+  static void Unlock() { gStateLock->Unlock(); }
+  static void Lock() { gStateLock->Lock(); }
+  static bool IsLocked() { return gStateLock->IsLocked(); }
+};
+
+typedef CodeAddressService<InfallibleAllocPolicy, DescribeCodeAddressLock>
+    CodeAddressService;
+
+//---------------------------------------------------------------------------
+// Stack traces
+//---------------------------------------------------------------------------
+
+class StackTrace {
+ public:
+  static const uint32_t MaxFrames = 24;
+
+ private:
+  uint32_t mLength;             // The number of PCs.
+  const void* mPcs[MaxFrames];  // The PCs themselves.
+
+ public:
+  StackTrace() : mLength(0) {}
+  StackTrace(const StackTrace& aOther) : mLength(aOther.mLength) {
+    PodCopy(mPcs, aOther.mPcs, mLength);
+  }
+
+  uint32_t Length() const { return mLength; }
+  const void* Pc(uint32_t i) const {
+    MOZ_ASSERT(i < mLength);
+    return mPcs[i];
+  }
+
+  uint32_t Size() const { return mLength * sizeof(mPcs[0]); }
+
+  // The stack trace returned by this function is interned in gStackTraceTable,
+  // and so is immortal and unmovable.
+  static const StackTrace* Get(Thread* aT);
+
+  // Hash policy.
+
+  typedef StackTrace* Lookup;
+
+  static mozilla::HashNumber hash(const StackTrace* const& aSt) {
+    return mozilla::HashBytes(aSt->mPcs, aSt->Size());
+  }
+
+  static bool match(const StackTrace* const& aA, const StackTrace* const& aB) {
+    return aA->mLength == aB->mLength &&
+           memcmp(aA->mPcs, aB->mPcs, aA->Size()) == 0;
+  }
+
+ private:
+  static void StackWalkCallback(uint32_t aFrameNumber, void* aPc, void* aSp,
+                                void* aClosure) {
+    StackTrace* st = (StackTrace*)aClosure;
+    MOZ_ASSERT(st->mLength < MaxFrames);
+    st->mPcs[st->mLength] = aPc;
+    st->mLength++;
+    MOZ_ASSERT(st->mLength == aFrameNumber);
+  }
+};
+
+typedef mozilla::HashSet<StackTrace*, StackTrace, InfallibleAllocPolicy>
+    StackTraceTable;
+static StackTraceTable* gStackTraceTable = nullptr;
+
+typedef mozilla::HashSet<const StackTrace*,
+                         mozilla::DefaultHasher<const StackTrace*>,
+                         InfallibleAllocPolicy>
+    StackTraceSet;
+
+typedef mozilla::HashSet<const void*, mozilla::DefaultHasher<const void*>,
+                         InfallibleAllocPolicy>
+    PointerSet;
+typedef mozilla::HashMap<const void*, uint32_t,
+                         mozilla::DefaultHasher<const void*>,
+                         InfallibleAllocPolicy>
+    PointerIdMap;
+
+// We won't GC the stack trace table until it this many elements.
+static uint32_t gGCStackTraceTableWhenSizeExceeds = 4 * 1024;
+
+/* static */ const StackTrace* StackTrace::Get(Thread* aT) {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(aT->InterceptsAreBlocked());
+
+  // On Windows, MozStackWalk can acquire a lock from the shared library
+  // loader.  Another thread might call malloc while holding that lock (when
+  // loading a shared library).  So we can't be in gStateLock during the call
+  // to MozStackWalk.  For details, see
+  // https://bugzilla.mozilla.org/show_bug.cgi?id=374829#c8
+  // On Linux, something similar can happen;  see bug 824340.
+  // So let's just release it on all platforms.
+  StackTrace tmp;
+  {
+    AutoUnlockState unlock;
+    // In each of the following cases, skipFrames is chosen so that the
+    // first frame in each stack trace is a replace_* function (or as close as
+    // possible, given the vagaries of inlining on different platforms).
+#if defined(XP_WIN) && defined(_M_IX86)
+    // This avoids MozStackWalk(), which causes unusably slow startup on Win32
+    // when it is called during static initialization (see bug 1241684).
+    //
+    // This code is cribbed from the Gecko Profiler, which also uses
+    // FramePointerStackWalk() on Win32: Registers::SyncPopulate() for the
+    // frame pointer, and GetStackTop() for the stack end.
+    CONTEXT context;
+    RtlCaptureContext(&context);
+    void** fp = reinterpret_cast<void**>(context.Ebp);
+
+    PNT_TIB pTib = reinterpret_cast<PNT_TIB>(NtCurrentTeb());
+    void* stackEnd = static_cast<void*>(pTib->StackBase);
+    FramePointerStackWalk(StackWalkCallback, /* skipFrames = */ 0, MaxFrames,
+                          &tmp, fp, stackEnd);
+#elif defined(XP_MACOSX)
+    // This avoids MozStackWalk(), which has become unusably slow on Mac due to
+    // changes in libunwind.
+    //
+    // This code is cribbed from the Gecko Profiler, which also uses
+    // FramePointerStackWalk() on Mac: Registers::SyncPopulate() for the frame
+    // pointer, and GetStackTop() for the stack end.
+    void** fp;
+#  if defined(__x86_64__)
+    asm(
+        // Dereference %rbp to get previous %rbp
+        "movq (%%rbp), %0\n\t"
+        : "=r"(fp));
+#  else
+    asm("ldr %0, [x29]\n\t" : "=r"(fp));
+#  endif
+    void* stackEnd = pthread_get_stackaddr_np(pthread_self());
+    FramePointerStackWalk(StackWalkCallback, /* skipFrames = */ 0, MaxFrames,
+                          &tmp, fp, stackEnd);
+#else
+#  if defined(XP_WIN) && defined(_M_X64)
+    int skipFrames = 1;
+#  else
+    int skipFrames = 2;
+#  endif
+    MozStackWalk(StackWalkCallback, skipFrames, MaxFrames, &tmp);
+#endif
+  }
+
+  StackTraceTable::AddPtr p = gStackTraceTable->lookupForAdd(&tmp);
+  if (!p) {
+    StackTrace* stnew = InfallibleAllocPolicy::new_<StackTrace>(tmp);
+    MOZ_ALWAYS_TRUE(gStackTraceTable->add(p, stnew));
+  }
+  return *p;
+}
+
+//---------------------------------------------------------------------------
+// Heap blocks
+//---------------------------------------------------------------------------
+
+// This class combines a 2-byte-aligned pointer (i.e. one whose bottom bit
+// is zero) with a 1-bit tag.
+//
+// |T| is the pointer type, e.g. |int*|, not the pointed-to type.  This makes
+// is easier to have const pointers, e.g. |TaggedPtr<const int*>|.
+template <typename T>
+class TaggedPtr {
+  union {
+    T mPtr;
+    uintptr_t mUint;
+  };
+
+  static const uintptr_t kTagMask = uintptr_t(0x1);
+  static const uintptr_t kPtrMask = ~kTagMask;
+
+  static bool IsTwoByteAligned(T aPtr) {
+    return (uintptr_t(aPtr) & kTagMask) == 0;
+  }
+
+ public:
+  TaggedPtr() : mPtr(nullptr) {}
+
+  TaggedPtr(T aPtr, bool aBool) : mPtr(aPtr) {
+    MOZ_ASSERT(IsTwoByteAligned(aPtr));
+    uintptr_t tag = uintptr_t(aBool);
+    MOZ_ASSERT(tag <= kTagMask);
+    mUint |= (tag & kTagMask);
+  }
+
+  void Set(T aPtr, bool aBool) {
+    MOZ_ASSERT(IsTwoByteAligned(aPtr));
+    mPtr = aPtr;
+    uintptr_t tag = uintptr_t(aBool);
+    MOZ_ASSERT(tag <= kTagMask);
+    mUint |= (tag & kTagMask);
+  }
+
+  T Ptr() const { return reinterpret_cast<T>(mUint & kPtrMask); }
+
+  bool Tag() const { return bool(mUint & kTagMask); }
+};
+
+// A live heap block. Stores both basic data and data about reports, if we're
+// in DarkMatter mode.
+class LiveBlock {
+  const void* mPtr;
+  const size_t mReqSize;  // size requested
+
+  // The stack trace where this block was allocated, or nullptr if we didn't
+  // record one.
+  const StackTrace* const mAllocStackTrace;
+
+  // This array has two elements because we record at most two reports of a
+  // block.
+  // - Ptr: |mReportStackTrace| - stack trace where this block was reported.
+  //   nullptr if not reported.
+  // - Tag bit 0: |mReportedOnAlloc| - was the block reported immediately on
+  //   allocation?  If so, DMD must not clear the report at the end of
+  //   Analyze(). Only relevant if |mReportStackTrace| is non-nullptr.
+  //
+  // |mPtr| is used as the key in LiveBlockTable, so it's ok for this member
+  // to be |mutable|.
+  //
+  // Only used in DarkMatter mode.
+  mutable TaggedPtr<const StackTrace*> mReportStackTrace_mReportedOnAlloc[2];
+
+ public:
+  LiveBlock(const void* aPtr, size_t aReqSize,
+            const StackTrace* aAllocStackTrace)
+      : mPtr(aPtr),
+        mReqSize(aReqSize),
+        mAllocStackTrace(aAllocStackTrace),
+        mReportStackTrace_mReportedOnAlloc()  // all fields get zeroed
+  {}
+
+  const void* Address() const { return mPtr; }
+
+  size_t ReqSize() const { return mReqSize; }
+
+  size_t SlopSize() const { return MallocSizeOf(mPtr) - mReqSize; }
+
+  const StackTrace* AllocStackTrace() const { return mAllocStackTrace; }
+
+  const StackTrace* ReportStackTrace1() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[0].Ptr();
+  }
+
+  const StackTrace* ReportStackTrace2() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[1].Ptr();
+  }
+
+  bool ReportedOnAlloc1() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[0].Tag();
+  }
+
+  bool ReportedOnAlloc2() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[1].Tag();
+  }
+
+  void AddStackTracesToTable(StackTraceSet& aStackTraces) const {
+    if (AllocStackTrace()) {
+      MOZ_ALWAYS_TRUE(aStackTraces.put(AllocStackTrace()));
+    }
+    if (gOptions->IsDarkMatterMode()) {
+      if (ReportStackTrace1()) {
+        MOZ_ALWAYS_TRUE(aStackTraces.put(ReportStackTrace1()));
+      }
+      if (ReportStackTrace2()) {
+        MOZ_ALWAYS_TRUE(aStackTraces.put(ReportStackTrace2()));
+      }
+    }
+  }
+
+  uint32_t NumReports() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    if (ReportStackTrace2()) {
+      MOZ_ASSERT(ReportStackTrace1());
+      return 2;
+    }
+    if (ReportStackTrace1()) {
+      return 1;
+    }
+    return 0;
+  }
+
+  // This is |const| thanks to the |mutable| fields above.
+  void Report(Thread* aT, bool aReportedOnAlloc) const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    // We don't bother recording reports after the 2nd one.
+    uint32_t numReports = NumReports();
+    if (numReports < 2) {
+      mReportStackTrace_mReportedOnAlloc[numReports].Set(StackTrace::Get(aT),
+                                                         aReportedOnAlloc);
+    }
+  }
+
+  void UnreportIfNotReportedOnAlloc() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    if (!ReportedOnAlloc1() && !ReportedOnAlloc2()) {
+      mReportStackTrace_mReportedOnAlloc[0].Set(nullptr, 0);
+      mReportStackTrace_mReportedOnAlloc[1].Set(nullptr, 0);
+
+    } else if (!ReportedOnAlloc1() && ReportedOnAlloc2()) {
+      // Shift the 2nd report down to the 1st one.
+      mReportStackTrace_mReportedOnAlloc[0] =
+          mReportStackTrace_mReportedOnAlloc[1];
+      mReportStackTrace_mReportedOnAlloc[1].Set(nullptr, 0);
+
+    } else if (ReportedOnAlloc1() && !ReportedOnAlloc2()) {
+      mReportStackTrace_mReportedOnAlloc[1].Set(nullptr, 0);
+    }
+  }
+
+  // Hash policy.
+
+  typedef const void* Lookup;
+
+  static mozilla::HashNumber hash(const void* const& aPtr) {
+    return mozilla::HashGeneric(aPtr);
+  }
+
+  static bool match(const LiveBlock& aB, const void* const& aPtr) {
+    return aB.mPtr == aPtr;
+  }
+};
+
+// A table of live blocks where the lookup key is the block address.
+typedef mozilla::HashSet<LiveBlock, LiveBlock, InfallibleAllocPolicy>
+    LiveBlockTable;
+static LiveBlockTable* gLiveBlockTable = nullptr;
+
+class AggregatedLiveBlockHashPolicy {
+ public:
+  typedef const LiveBlock* const Lookup;
+
+  static mozilla::HashNumber hash(const LiveBlock* const& aB) {
+    return gOptions->IsDarkMatterMode()
+               ? mozilla::HashGeneric(
+                     aB->ReqSize(), aB->SlopSize(), aB->AllocStackTrace(),
+                     aB->ReportedOnAlloc1(), aB->ReportedOnAlloc2())
+               : mozilla::HashGeneric(aB->ReqSize(), aB->SlopSize(),
+                                      aB->AllocStackTrace());
+  }
+
+  static bool match(const LiveBlock* const& aA, const LiveBlock* const& aB) {
+    return gOptions->IsDarkMatterMode()
+               ? aA->ReqSize() == aB->ReqSize() &&
+                     aA->SlopSize() == aB->SlopSize() &&
+                     aA->AllocStackTrace() == aB->AllocStackTrace() &&
+                     aA->ReportStackTrace1() == aB->ReportStackTrace1() &&
+                     aA->ReportStackTrace2() == aB->ReportStackTrace2()
+               : aA->ReqSize() == aB->ReqSize() &&
+                     aA->SlopSize() == aB->SlopSize() &&
+                     aA->AllocStackTrace() == aB->AllocStackTrace();
+  }
+};
+
+// A table of live blocks where the lookup key is everything but the block
+// address. For aggregating similar live blocks at output time.
+typedef mozilla::HashMap<const LiveBlock*, size_t,
+                         AggregatedLiveBlockHashPolicy, InfallibleAllocPolicy>
+    AggregatedLiveBlockTable;
+
+// A freed heap block.
+class DeadBlock {
+  const size_t mReqSize;   // size requested
+  const size_t mSlopSize;  // slop above size requested
+
+  // The stack trace where this block was allocated.
+  const StackTrace* const mAllocStackTrace;
+
+ public:
+  DeadBlock() : mReqSize(0), mSlopSize(0), mAllocStackTrace(nullptr) {}
+
+  explicit DeadBlock(const LiveBlock& aLb)
+      : mReqSize(aLb.ReqSize()),
+        mSlopSize(aLb.SlopSize()),
+        mAllocStackTrace(aLb.AllocStackTrace()) {}
+
+  ~DeadBlock() {}
+
+  size_t ReqSize() const { return mReqSize; }
+  size_t SlopSize() const { return mSlopSize; }
+
+  const StackTrace* AllocStackTrace() const { return mAllocStackTrace; }
+
+  void AddStackTracesToTable(StackTraceSet& aStackTraces) const {
+    if (AllocStackTrace()) {
+      MOZ_ALWAYS_TRUE(aStackTraces.put(AllocStackTrace()));
+    }
+  }
+
+  // Hash policy.
+
+  typedef DeadBlock Lookup;
+
+  static mozilla::HashNumber hash(const DeadBlock& aB) {
+    return mozilla::HashGeneric(aB.ReqSize(), aB.SlopSize(),
+                                aB.AllocStackTrace());
+  }
+
+  static bool match(const DeadBlock& aA, const DeadBlock& aB) {
+    return aA.ReqSize() == aB.ReqSize() && aA.SlopSize() == aB.SlopSize() &&
+           aA.AllocStackTrace() == aB.AllocStackTrace();
+  }
+};
+
+// For each unique DeadBlock value we store a count of how many actual dead
+// blocks have that value.
+typedef mozilla::HashMap<DeadBlock, size_t, DeadBlock, InfallibleAllocPolicy>
+    DeadBlockTable;
+static DeadBlockTable* gDeadBlockTable = nullptr;
+
+// Add the dead block to the dead block table, if that's appropriate.
+void MaybeAddToDeadBlockTable(const DeadBlock& aDb) {
+  if (gOptions->IsCumulativeMode() && aDb.AllocStackTrace()) {
+    AutoLockState lock;
+    if (DeadBlockTable::AddPtr p = gDeadBlockTable->lookupForAdd(aDb)) {
+      p->value() += 1;
+    } else {
+      MOZ_ALWAYS_TRUE(gDeadBlockTable->add(p, aDb, 1));
+    }
+  }
+}
+
+// Add a pointer to each live stack trace into the given StackTraceSet.  (A
+// stack trace is live if it's used by one of the live blocks.)
+static void GatherUsedStackTraces(StackTraceSet& aStackTraces) {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(Thread::Fetch()->InterceptsAreBlocked());
+
+  aStackTraces.clear();
+  MOZ_ALWAYS_TRUE(aStackTraces.reserve(512));
+
+  for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+    iter.get().AddStackTracesToTable(aStackTraces);
+  }
+
+  for (auto iter = gDeadBlockTable->iter(); !iter.done(); iter.next()) {
+    iter.get().key().AddStackTracesToTable(aStackTraces);
+  }
+}
+
+// Delete stack traces that we aren't using, and compact our hashtable.
+static void GCStackTraces() {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(Thread::Fetch()->InterceptsAreBlocked());
+
+  StackTraceSet usedStackTraces;
+  GatherUsedStackTraces(usedStackTraces);
+
+  // Delete all unused stack traces from gStackTraceTable.  The ModIterator
+  // destructor will automatically rehash and compact the table.
+  for (auto iter = gStackTraceTable->modIter(); !iter.done(); iter.next()) {
+    StackTrace* const& st = iter.get();
+    if (!usedStackTraces.has(st)) {
+      iter.remove();
+      InfallibleAllocPolicy::delete_(st);
+    }
+  }
+
+  // Schedule a GC when we have twice as many stack traces as we had right after
+  // this GC finished.
+  gGCStackTraceTableWhenSizeExceeds = 2 * gStackTraceTable->count();
+}
+
+//---------------------------------------------------------------------------
+// malloc/free callbacks
+//---------------------------------------------------------------------------
+
+static FastBernoulliTrial* gBernoulli;
+
+// In testing, a probability of 0.003 resulted in ~25% of heap blocks getting
+// a stack trace and ~80% of heap bytes getting a stack trace. (This is
+// possible because big heap blocks are more likely to get a stack trace.)
+//
+// We deliberately choose not to give the user control over this probability
+// (other than effectively setting it to 1 via --stacks=full) because it's
+// quite inscrutable and generally the user just wants "faster and imprecise"
+// or "slower and precise".
+//
+// The random number seeds are arbitrary and were obtained from random.org. If
+// you change them you'll need to change the tests as well, because their
+// expected output is based on the particular sequence of trial results that we
+// get with these seeds.
+static void ResetBernoulli() {
+  new (gBernoulli)
+      FastBernoulliTrial(0.003, 0x8e26eeee166bc8ca, 0x56820f304a9c9ae0);
+}
+
+static void AllocCallback(void* aPtr, size_t aReqSize, Thread* aT) {
+  if (!aPtr) {
+    return;
+  }
+
+  AutoLockState lock;
+  AutoBlockIntercepts block(aT);
+
+  size_t actualSize = gMallocTable.malloc_usable_size(aPtr);
+
+  // We may or may not record the allocation stack trace, depending on the
+  // options and the outcome of a Bernoulli trial.
+  bool getTrace = gOptions->DoFullStacks() || gBernoulli->trial(actualSize);
+  LiveBlock b(aPtr, aReqSize, getTrace ? StackTrace::Get(aT) : nullptr);
+  MOZ_ALWAYS_TRUE(gLiveBlockTable->putNew(aPtr, b));
+}
+
+static void FreeCallback(void* aPtr, Thread* aT, DeadBlock* aDeadBlock) {
+  if (!aPtr) {
+    return;
+  }
+
+  AutoLockState lock;
+  AutoBlockIntercepts block(aT);
+
+  if (LiveBlockTable::Ptr lb = gLiveBlockTable->lookup(aPtr)) {
+    if (gOptions->IsCumulativeMode()) {
+      // Copy it out so it can be added to the dead block list later.
+      new (aDeadBlock) DeadBlock(*lb);
+    }
+    gLiveBlockTable->remove(lb);
+  } else {
+    // We have no record of the block. It must be a bogus pointer, or one that
+    // DMD wasn't able to see allocated. This should be extremely rare.
+  }
+
+  if (gStackTraceTable->count() > gGCStackTraceTableWhenSizeExceeds) {
+    GCStackTraces();
+  }
+}
+
+//---------------------------------------------------------------------------
+// malloc/free interception
+//---------------------------------------------------------------------------
+
+static bool Init(malloc_table_t* aMallocTable);
+
+}  // namespace dmd
+}  // namespace mozilla
+
+static void* replace_malloc(size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    // Intercepts are blocked, which means this must be a call to malloc
+    // triggered indirectly by DMD (e.g. via MozStackWalk).  Be infallible.
+    return InfallibleAllocPolicy::malloc_(aSize);
+  }
+
+  // This must be a call to malloc from outside DMD.  Intercept it.
+  void* ptr = gMallocTable.malloc(aSize);
+  AllocCallback(ptr, aSize, t);
+  return ptr;
+}
+
+static void* replace_calloc(size_t aCount, size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::calloc_(aCount, aSize);
+  }
+
+  // |aCount * aSize| could overflow, but if that happens then
+  // |gMallocTable.calloc()| will return nullptr and |AllocCallback()| will
+  // return immediately without using the overflowed value.
+  void* ptr = gMallocTable.calloc(aCount, aSize);
+  AllocCallback(ptr, aCount * aSize, t);
+  return ptr;
+}
+
+static void* replace_realloc(void* aOldPtr, size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::realloc_(aOldPtr, aSize);
+  }
+
+  // If |aOldPtr| is nullptr, the call is equivalent to |malloc(aSize)|.
+  if (!aOldPtr) {
+    return replace_malloc(aSize);
+  }
+
+  // Be very careful here!  Must remove the block from the table before doing
+  // the realloc to avoid races, just like in replace_free().
+  // Nb: This does an unnecessary hashtable remove+add if the block doesn't
+  // move, but doing better isn't worth the effort.
+  DeadBlock db;
+  FreeCallback(aOldPtr, t, &db);
+  void* ptr = gMallocTable.realloc(aOldPtr, aSize);
+  if (ptr) {
+    AllocCallback(ptr, aSize, t);
+    MaybeAddToDeadBlockTable(db);
+  } else {
+    // If realloc fails, we undo the prior operations by re-inserting the old
+    // pointer into the live block table. We don't have to do anything with the
+    // dead block list because the dead block hasn't yet been inserted. The
+    // block will end up looking like it was allocated for the first time here,
+    // which is untrue, and the slop bytes will be zero, which may be untrue.
+    // But this case is rare and doing better isn't worth the effort.
+    AllocCallback(aOldPtr, gMallocTable.malloc_usable_size(aOldPtr), t);
+  }
+  return ptr;
+}
+
+static void* replace_memalign(size_t aAlignment, size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::memalign_(aAlignment, aSize);
+  }
+
+  void* ptr = gMallocTable.memalign(aAlignment, aSize);
+  AllocCallback(ptr, aSize, t);
+  return ptr;
+}
+
+static void replace_free(void* aPtr) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::free_(aPtr);
+  }
+
+  // Do the actual free after updating the table.  Otherwise, another thread
+  // could call malloc and get the freed block and update the table, and then
+  // our update here would remove the newly-malloc'd block.
+  DeadBlock db;
+  FreeCallback(aPtr, t, &db);
+  MaybeAddToDeadBlockTable(db);
+  gMallocTable.free(aPtr);
+}
+
+void replace_init(malloc_table_t* aMallocTable, ReplaceMallocBridge** aBridge) {
+  if (mozilla::dmd::Init(aMallocTable)) {
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#define MALLOC_DECL(name, ...) aMallocTable->name = replace_##name;
+#include "malloc_decls.h"
+    *aBridge = mozilla::dmd::gDMDBridge;
+  }
+}
+
+namespace mozilla {
+namespace dmd {
+
+//---------------------------------------------------------------------------
+// Options (Part 2)
+//---------------------------------------------------------------------------
+
+// Given an |aOptionName| like "foo", succeed if |aArg| has the form "foo=blah"
+// (where "blah" is non-empty) and return the pointer to "blah".  |aArg| can
+// have leading space chars (but not other whitespace).
+const char* Options::ValueIfMatch(const char* aArg, const char* aOptionName) {
+  MOZ_ASSERT(!isspace(*aArg));  // any leading whitespace should not remain
+  size_t optionLen = strlen(aOptionName);
+  if (strncmp(aArg, aOptionName, optionLen) == 0 && aArg[optionLen] == '=' &&
+      aArg[optionLen + 1]) {
+    return aArg + optionLen + 1;
+  }
+  return nullptr;
+}
+
+// Extracts a |long| value for an option from an argument.  It must be within
+// the range |aMin..aMax| (inclusive).
+bool Options::GetLong(const char* aArg, const char* aOptionName, long aMin,
+                      long aMax, long* aValue) {
+  if (const char* optionValue = ValueIfMatch(aArg, aOptionName)) {
+    char* endPtr;
+    *aValue = strtol(optionValue, &endPtr, /* base */ 10);
+    if (!*endPtr && aMin <= *aValue && *aValue <= aMax && *aValue != LONG_MIN &&
+        *aValue != LONG_MAX) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Extracts a |bool| value for an option -- encoded as "yes" or "no" -- from an
+// argument.
+bool Options::GetBool(const char* aArg, const char* aOptionName, bool* aValue) {
+  if (const char* optionValue = ValueIfMatch(aArg, aOptionName)) {
+    if (strcmp(optionValue, "yes") == 0) {
+      *aValue = true;
+      return true;
+    }
+    if (strcmp(optionValue, "no") == 0) {
+      *aValue = false;
+      return true;
+    }
+  }
+  return false;
+}
+
+Options::Options(const char* aDMDEnvVar)
+    : mDMDEnvVar(aDMDEnvVar ? InfallibleAllocPolicy::strdup_(aDMDEnvVar)
+                            : nullptr),
+      mMode(Mode::DarkMatter),
+      mStacks(Stacks::Partial),
+      mShowDumpStats(false) {
+  char* e = mDMDEnvVar;
+  if (e && strcmp(e, "1") != 0) {
+    bool isEnd = false;
+    while (!isEnd) {
+      // Consume leading whitespace.
+      while (isspace(*e)) {
+        e++;
+      }
+
+      // Save the start of the arg.
+      const char* arg = e;
+
+      // Find the first char after the arg, and temporarily change it to '\0'
+      // to isolate the arg.
+      while (!isspace(*e) && *e != '\0') {
+        e++;
+      }
+      char replacedChar = *e;
+      isEnd = replacedChar == '\0';
+      *e = '\0';
+
+      // Handle arg
+      bool myBool;
+      if (strcmp(arg, "--mode=live") == 0) {
+        mMode = Mode::Live;
+      } else if (strcmp(arg, "--mode=dark-matter") == 0) {
+        mMode = Mode::DarkMatter;
+      } else if (strcmp(arg, "--mode=cumulative") == 0) {
+        mMode = Mode::Cumulative;
+      } else if (strcmp(arg, "--mode=scan") == 0) {
+        mMode = Mode::Scan;
+
+      } else if (strcmp(arg, "--stacks=full") == 0) {
+        mStacks = Stacks::Full;
+      } else if (strcmp(arg, "--stacks=partial") == 0) {
+        mStacks = Stacks::Partial;
+
+      } else if (GetBool(arg, "--show-dump-stats", &myBool)) {
+        mShowDumpStats = myBool;
+
+      } else if (strcmp(arg, "") == 0) {
+        // This can only happen if there is trailing whitespace.  Ignore.
+        MOZ_ASSERT(isEnd);
+
+      } else {
+        BadArg(arg);
+      }
+
+      // Undo the temporary isolation.
+      *e = replacedChar;
+    }
+  }
+
+  if (mMode == Mode::Scan) {
+    mStacks = Stacks::Full;
+  }
+}
+
+void Options::BadArg(const char* aArg) {
+  StatusMsg("\n");
+  StatusMsg("Bad entry in the $DMD environment variable: '%s'.\n", aArg);
+  StatusMsg("See the output of |mach help run| for the allowed options.\n");
+  exit(1);
+}
+
+const char* Options::ModeString() const {
+  switch (mMode) {
+    case Mode::Live:
+      return "live";
+    case Mode::DarkMatter:
+      return "dark-matter";
+    case Mode::Cumulative:
+      return "cumulative";
+    case Mode::Scan:
+      return "scan";
+    default:
+      MOZ_ASSERT(false);
+      return "(unknown DMD mode)";
+  }
+}
+
+//---------------------------------------------------------------------------
+// DMD start-up
+//---------------------------------------------------------------------------
+
+#ifndef XP_WIN
+static void prefork() {
+  if (gStateLock) {
+    gStateLock->Lock();
+  }
+}
+
+static void postfork() {
+  if (gStateLock) {
+    gStateLock->Unlock();
+  }
+}
+#endif
+
+// WARNING: this function runs *very* early -- before all static initializers
+// have run.  For this reason, non-scalar globals such as gStateLock and
+// gStackTraceTable are allocated dynamically (so we can guarantee their
+// construction in this function) rather than statically.
+static bool Init(malloc_table_t* aMallocTable) {
+  // DMD is controlled by the |DMD| environment variable.
+  const char* e = getenv("DMD");
+
+  if (!e) {
+    return false;
+  }
+  // Initialize the function table first, because StatusMsg uses
+  // InfallibleAllocPolicy::malloc_, which uses it.
+  gMallocTable = *aMallocTable;
+
+  StatusMsg("$DMD = '%s'\n", e);
+
+  gDMDBridge = InfallibleAllocPolicy::new_<DMDBridge>();
+
+#ifndef XP_WIN
+  // Avoid deadlocks when forking by acquiring our state lock prior to forking
+  // and releasing it after forking. See |LogAlloc|'s |replace_init| for
+  // in-depth details.
+  //
+  // Note: This must run after attempting an allocation so as to give the
+  // system malloc a chance to insert its own atfork handler.
+  pthread_atfork(prefork, postfork, postfork);
+#endif
+  // Parse $DMD env var.
+  gOptions = InfallibleAllocPolicy::new_<Options>(e);
+
+  gStateLock = InfallibleAllocPolicy::new_<Mutex>();
+
+  gBernoulli = (FastBernoulliTrial*)InfallibleAllocPolicy::malloc_(
+      sizeof(FastBernoulliTrial));
+  ResetBernoulli();
+
+  Thread::Init();
+
+  {
+    AutoLockState lock;
+
+    gStackTraceTable = InfallibleAllocPolicy::new_<StackTraceTable>(8192);
+    gLiveBlockTable = InfallibleAllocPolicy::new_<LiveBlockTable>(8192);
+
+    // Create this even if the mode isn't Cumulative (albeit with a small
+    // size), in case the mode is changed later on (as is done by SmokeDMD.cpp,
+    // for example).
+    size_t tableSize = gOptions->IsCumulativeMode() ? 8192 : 4;
+    gDeadBlockTable = InfallibleAllocPolicy::new_<DeadBlockTable>(tableSize);
+  }
+
+  return true;
+}
+
+//---------------------------------------------------------------------------
+// Block reporting and unreporting
+//---------------------------------------------------------------------------
+
+static void ReportHelper(const void* aPtr, bool aReportedOnAlloc) {
+  if (!gOptions->IsDarkMatterMode() || !aPtr) {
+    return;
+  }
+
+  Thread* t = Thread::Fetch();
+
+  AutoBlockIntercepts block(t);
+  AutoLockState lock;
+
+  if (LiveBlockTable::Ptr p = gLiveBlockTable->lookup(aPtr)) {
+    p->Report(t, aReportedOnAlloc);
+  } else {
+    // We have no record of the block. It must be a bogus pointer. This should
+    // be extremely rare because Report() is almost always called in
+    // conjunction with a malloc_size_of-style function. Print a message so
+    // that we get some feedback.
+    StatusMsg("Unknown pointer %p\n", aPtr);
+  }
+}
+
+void DMDFuncs::Report(const void* aPtr) {
+  ReportHelper(aPtr, /* onAlloc */ false);
+}
+
+void DMDFuncs::ReportOnAlloc(const void* aPtr) {
+  ReportHelper(aPtr, /* onAlloc */ true);
+}
+
+//---------------------------------------------------------------------------
+// DMD output
+//---------------------------------------------------------------------------
+
+// The version number of the output format. Increment this if you make
+// backwards-incompatible changes to the format. See DMD.h for the version
+// history.
+static const int kOutputVersionNumber = 5;
+
+// Note that, unlike most SizeOf* functions, this function does not take a
+// |mozilla::MallocSizeOf| argument.  That's because those arguments are
+// primarily to aid DMD track heap blocks... but DMD deliberately doesn't track
+// heap blocks it allocated for itself!
+//
+// SizeOfInternal should be called while you're holding the state lock and
+// while intercepts are blocked; SizeOf acquires the lock and blocks
+// intercepts.
+
+static void SizeOfInternal(Sizes* aSizes) {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(Thread::Fetch()->InterceptsAreBlocked());
+
+  aSizes->Clear();
+
+  StackTraceSet usedStackTraces;
+  GatherUsedStackTraces(usedStackTraces);
+
+  for (auto iter = gStackTraceTable->iter(); !iter.done(); iter.next()) {
+    StackTrace* const& st = iter.get();
+
+    if (usedStackTraces.has(st)) {
+      aSizes->mStackTracesUsed += MallocSizeOf(st);
+    } else {
+      aSizes->mStackTracesUnused += MallocSizeOf(st);
+    }
+  }
+
+  aSizes->mStackTraceTable =
+      gStackTraceTable->shallowSizeOfIncludingThis(MallocSizeOf);
+
+  aSizes->mLiveBlockTable =
+      gLiveBlockTable->shallowSizeOfIncludingThis(MallocSizeOf);
+
+  aSizes->mDeadBlockTable =
+      gDeadBlockTable->shallowSizeOfIncludingThis(MallocSizeOf);
+}
+
+void DMDFuncs::SizeOf(Sizes* aSizes) {
+  aSizes->Clear();
+
+  AutoBlockIntercepts block(Thread::Fetch());
+  AutoLockState lock;
+  SizeOfInternal(aSizes);
+}
+
+void DMDFuncs::ClearReports() {
+  if (!gOptions->IsDarkMatterMode()) {
+    return;
+  }
+
+  AutoLockState lock;
+
+  // Unreport all blocks that were marked reported by a memory reporter.  This
+  // excludes those that were reported on allocation, because they need to keep
+  // their reported marking.
+  for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+    iter.get().UnreportIfNotReportedOnAlloc();
+  }
+}
+
+class ToIdStringConverter final {
+ public:
+  ToIdStringConverter() : mIdMap(512), mNextId(0) {}
+
+  // Converts a pointer to a unique ID. Reuses the existing ID for the pointer
+  // if it's been seen before.
+  const char* ToIdString(const void* aPtr) {
+    uint32_t id;
+    PointerIdMap::AddPtr p = mIdMap.lookupForAdd(aPtr);
+    if (!p) {
+      id = mNextId++;
+      MOZ_ALWAYS_TRUE(mIdMap.add(p, aPtr, id));
+    } else {
+      id = p->value();
+    }
+    return Base32(id);
+  }
+
+  size_t sizeOfExcludingThis(mozilla::MallocSizeOf aMallocSizeOf) const {
+    return mIdMap.shallowSizeOfExcludingThis(aMallocSizeOf);
+  }
+
+ private:
+  // This function converts an integer to base-32. We use base-32 values for
+  // indexing into the traceTable and the frameTable, for the following reasons.
+  //
+  // - Base-32 gives more compact indices than base-16.
+  //
+  // - 32 is a power-of-two, which makes the necessary div/mod calculations
+  //   fast.
+  //
+  // - We can (and do) choose non-numeric digits for base-32. When
+  //   inspecting/debugging the JSON output, non-numeric indices are easier to
+  //   search for than numeric indices.
+  //
+  char* Base32(uint32_t aN) {
+    static const char digits[] = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdef";
+
+    char* b = mIdBuf + kIdBufLen - 1;
+    *b = '\0';
+    do {
+      b--;
+      if (b == mIdBuf) {
+        MOZ_CRASH("Base32 buffer too small");
+      }
+      *b = digits[aN % 32];
+      aN /= 32;
+    } while (aN);
+
+    return b;
+  }
+
+  PointerIdMap mIdMap;
+  uint32_t mNextId;
+
+  // |mIdBuf| must have space for at least eight chars, which is the space
+  // needed to hold 'Dffffff' (including the terminating null char), which is
+  // the base-32 representation of 0xffffffff.
+  static const size_t kIdBufLen = 16;
+  char mIdBuf[kIdBufLen];
+};
+
+// Helper class for converting a pointer value to a string.
+class ToStringConverter {
+ public:
+  const char* ToPtrString(const void* aPtr) {
+    snprintf(kPtrBuf, sizeof(kPtrBuf) - 1, "%" PRIxPTR, (uintptr_t)aPtr);
+    return kPtrBuf;
+  }
+
+ private:
+  char kPtrBuf[32];
+};
+
+static void WriteBlockContents(JSONWriter& aWriter, const LiveBlock& aBlock) {
+  size_t numWords = aBlock.ReqSize() / sizeof(uintptr_t*);
+  if (numWords == 0) {
+    return;
+  }
+
+  aWriter.StartArrayProperty("contents", aWriter.SingleLineStyle);
+  {
+    const uintptr_t** block = (const uintptr_t**)aBlock.Address();
+    ToStringConverter sc;
+    for (size_t i = 0; i < numWords; ++i) {
+      aWriter.StringElement(MakeStringSpan(sc.ToPtrString(block[i])));
+    }
+  }
+  aWriter.EndArray();
+}
+
+static void AnalyzeImpl(UniquePtr<JSONWriteFunc> aWriter) {
+  // Some blocks may have been allocated while creating |aWriter|. Those blocks
+  // will be freed at the end of this function when |write| is destroyed. The
+  // allocations will have occurred while intercepts were not blocked, so the
+  // frees better be as well, otherwise we'll get assertion failures.
+  // Therefore, this declaration must precede the AutoBlockIntercepts
+  // declaration, to ensure that |write| is destroyed *after* intercepts are
+  // unblocked.
+  JSONWriter writer(std::move(aWriter));
+
+  AutoBlockIntercepts block(Thread::Fetch());
+  AutoLockState lock;
+
+  // Allocate this on the heap instead of the stack because it's fairly large.
+  auto locService = InfallibleAllocPolicy::new_<CodeAddressService>();
+
+  StackTraceSet usedStackTraces(512);
+  PointerSet usedPcs(512);
+
+  size_t iscSize;
+
+  static int analysisCount = 1;
+  StatusMsg("Dump %d {\n", analysisCount++);
+
+  writer.Start();
+  {
+    writer.IntProperty("version", kOutputVersionNumber);
+
+    writer.StartObjectProperty("invocation");
+    {
+      const char* var = gOptions->DMDEnvVar();
+      if (var) {
+        writer.StringProperty("dmdEnvVar", MakeStringSpan(var));
+      } else {
+        writer.NullProperty("dmdEnvVar");
+      }
+
+      writer.StringProperty("mode", MakeStringSpan(gOptions->ModeString()));
+    }
+    writer.EndObject();
+
+    StatusMsg("  Constructing the heap block list...\n");
+
+    ToIdStringConverter isc;
+    ToStringConverter sc;
+
+    writer.StartArrayProperty("blockList");
+    {
+      // Lambda that writes out a live block.
+      auto writeLiveBlock = [&](const LiveBlock& aB, size_t aNum) {
+        aB.AddStackTracesToTable(usedStackTraces);
+
+        MOZ_ASSERT_IF(gOptions->IsScanMode(), aNum == 1);
+
+        writer.StartObjectElement(writer.SingleLineStyle);
+        {
+          if (gOptions->IsScanMode()) {
+            writer.StringProperty("addr",
+                                  MakeStringSpan(sc.ToPtrString(aB.Address())));
+            WriteBlockContents(writer, aB);
+          }
+          writer.IntProperty("req", aB.ReqSize());
+          if (aB.SlopSize() > 0) {
+            writer.IntProperty("slop", aB.SlopSize());
+          }
+
+          if (aB.AllocStackTrace()) {
+            writer.StringProperty(
+                "alloc", MakeStringSpan(isc.ToIdString(aB.AllocStackTrace())));
+          }
+
+          if (gOptions->IsDarkMatterMode() && aB.NumReports() > 0) {
+            writer.StartArrayProperty("reps");
+            {
+              if (aB.ReportStackTrace1()) {
+                writer.StringElement(
+                    MakeStringSpan(isc.ToIdString(aB.ReportStackTrace1())));
+              }
+              if (aB.ReportStackTrace2()) {
+                writer.StringElement(
+                    MakeStringSpan(isc.ToIdString(aB.ReportStackTrace2())));
+              }
+            }
+            writer.EndArray();
+          }
+
+          if (aNum > 1) {
+            writer.IntProperty("num", aNum);
+          }
+        }
+        writer.EndObject();
+      };
+
+      // Live blocks.
+      if (!gOptions->IsScanMode()) {
+        // At this point we typically have many LiveBlocks that differ only in
+        // their address. Aggregate them to reduce the size of the output file.
+        AggregatedLiveBlockTable agg(8192);
+        for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+          const LiveBlock& b = iter.get();
+          b.AddStackTracesToTable(usedStackTraces);
+
+          if (AggregatedLiveBlockTable::AddPtr p = agg.lookupForAdd(&b)) {
+            p->value() += 1;
+          } else {
+            MOZ_ALWAYS_TRUE(agg.add(p, &b, 1));
+          }
+        }
+
+        // Now iterate over the aggregated table.
+        for (auto iter = agg.iter(); !iter.done(); iter.next()) {
+          const LiveBlock& b = *iter.get().key();
+          size_t num = iter.get().value();
+          writeLiveBlock(b, num);
+        }
+
+      } else {
+        // In scan mode we cannot aggregate because we print each live block's
+        // address and contents.
+        for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+          const LiveBlock& b = iter.get();
+          b.AddStackTracesToTable(usedStackTraces);
+
+          writeLiveBlock(b, 1);
+        }
+      }
+
+      // Dead blocks.
+      for (auto iter = gDeadBlockTable->iter(); !iter.done(); iter.next()) {
+        const DeadBlock& b = iter.get().key();
+        b.AddStackTracesToTable(usedStackTraces);
+
+        size_t num = iter.get().value();
+        MOZ_ASSERT(num > 0);
+
+        writer.StartObjectElement(writer.SingleLineStyle);
+        {
+          writer.IntProperty("req", b.ReqSize());
+          if (b.SlopSize() > 0) {
+            writer.IntProperty("slop", b.SlopSize());
+          }
+          if (b.AllocStackTrace()) {
+            writer.StringProperty(
+                "alloc", MakeStringSpan(isc.ToIdString(b.AllocStackTrace())));
+          }
+
+          if (num > 1) {
+            writer.IntProperty("num", num);
+          }
+        }
+        writer.EndObject();
+      }
+    }
+    writer.EndArray();
+
+    StatusMsg("  Constructing the stack trace table...\n");
+
+    writer.StartObjectProperty("traceTable");
+    {
+      for (auto iter = usedStackTraces.iter(); !iter.done(); iter.next()) {
+        const StackTrace* const st = iter.get();
+        writer.StartArrayProperty(MakeStringSpan(isc.ToIdString(st)),
+                                  writer.SingleLineStyle);
+        {
+          for (uint32_t i = 0; i < st->Length(); i++) {
+            const void* pc = st->Pc(i);
+            writer.StringElement(MakeStringSpan(isc.ToIdString(pc)));
+            MOZ_ALWAYS_TRUE(usedPcs.put(pc));
+          }
+        }
+        writer.EndArray();
+      }
+    }
+    writer.EndObject();
+
+    StatusMsg("  Constructing the stack frame table...\n");
+
+    writer.StartObjectProperty("frameTable");
+    {
+      static const size_t locBufLen = 1024;
+      char locBuf[locBufLen];
+
+      for (auto iter = usedPcs.iter(); !iter.done(); iter.next()) {
+        const void* const pc = iter.get();
+
+        // Use 0 for the frame number. See the JSON format description comment
+        // in DMD.h to understand why.
+        locService->GetLocation(0, pc, locBuf, locBufLen);
+        writer.StringProperty(MakeStringSpan(isc.ToIdString(pc)),
+                              MakeStringSpan(locBuf));
+      }
+    }
+    writer.EndObject();
+
+    iscSize = isc.sizeOfExcludingThis(MallocSizeOf);
+  }
+  writer.End();
+
+  if (gOptions->ShowDumpStats()) {
+    Sizes sizes;
+    SizeOfInternal(&sizes);
+
+    static const size_t kBufLen = 64;
+    char buf1[kBufLen];
+    char buf2[kBufLen];
+    char buf3[kBufLen];
+
+    StatusMsg("  Execution measurements {\n");
+
+    StatusMsg("    Data structures that persist after Dump() ends {\n");
+
+    StatusMsg("      Used stack traces:    %10s bytes\n",
+              Show(sizes.mStackTracesUsed, buf1, kBufLen));
+
+    StatusMsg("      Unused stack traces:  %10s bytes\n",
+              Show(sizes.mStackTracesUnused, buf1, kBufLen));
+
+    StatusMsg("      Stack trace table:    %10s bytes (%s entries, %s used)\n",
+              Show(sizes.mStackTraceTable, buf1, kBufLen),
+              Show(gStackTraceTable->capacity(), buf2, kBufLen),
+              Show(gStackTraceTable->count(), buf3, kBufLen));
+
+    StatusMsg("      Live block table:     %10s bytes (%s entries, %s used)\n",
+              Show(sizes.mLiveBlockTable, buf1, kBufLen),
+              Show(gLiveBlockTable->capacity(), buf2, kBufLen),
+              Show(gLiveBlockTable->count(), buf3, kBufLen));
+
+    StatusMsg("      Dead block table:     %10s bytes (%s entries, %s used)\n",
+              Show(sizes.mDeadBlockTable, buf1, kBufLen),
+              Show(gDeadBlockTable->capacity(), buf2, kBufLen),
+              Show(gDeadBlockTable->count(), buf3, kBufLen));
+
+    StatusMsg("    }\n");
+    StatusMsg("    Data structures that are destroyed after Dump() ends {\n");
+
+    StatusMsg(
+        "      Location service:      %10s bytes\n",
+        Show(locService->SizeOfIncludingThis(MallocSizeOf), buf1, kBufLen));
+    StatusMsg("      Used stack traces set: %10s bytes\n",
+              Show(usedStackTraces.shallowSizeOfExcludingThis(MallocSizeOf),
+                   buf1, kBufLen));
+    StatusMsg(
+        "      Used PCs set:          %10s bytes\n",
+        Show(usedPcs.shallowSizeOfExcludingThis(MallocSizeOf), buf1, kBufLen));
+    StatusMsg("      Pointer ID map:        %10s bytes\n",
+              Show(iscSize, buf1, kBufLen));
+
+    StatusMsg("    }\n");
+    StatusMsg("    Counts {\n");
+
+    size_t hits = locService->NumCacheHits();
+    size_t misses = locService->NumCacheMisses();
+    size_t requests = hits + misses;
+    StatusMsg("      Location service:    %10s requests\n",
+              Show(requests, buf1, kBufLen));
+
+    size_t count = locService->CacheCount();
+    size_t capacity = locService->CacheCapacity();
+    StatusMsg(
+        "      Location service cache:  "
+        "%4.1f%% hit rate, %.1f%% occupancy at end\n",
+        Percent(hits, requests), Percent(count, capacity));
+
+    StatusMsg("    }\n");
+    StatusMsg("  }\n");
+  }
+
+  InfallibleAllocPolicy::delete_(locService);
+
+  StatusMsg("}\n");
+}
+
+void DMDFuncs::Analyze(UniquePtr<JSONWriteFunc> aWriter) {
+  AnalyzeImpl(std::move(aWriter));
+  ClearReports();
+}
+
+//---------------------------------------------------------------------------
+// Testing
+//---------------------------------------------------------------------------
+
+void DMDFuncs::ResetEverything(const char* aOptions) {
+  AutoLockState lock;
+
+  // Reset options.
+  InfallibleAllocPolicy::delete_(gOptions);
+  gOptions = InfallibleAllocPolicy::new_<Options>(aOptions);
+
+  // Clear all existing blocks.
+  gLiveBlockTable->clear();
+  gDeadBlockTable->clear();
+
+  // Reset gBernoulli to a deterministic state. (Its current state depends on
+  // all previous trials.)
+  ResetBernoulli();
+}
+
+}  // namespace dmd
+}  // namespace mozilla
diff --git a/memory/replace/dmd/DMD.h b/memory/replace/dmd/DMD.h
new file mode 100644
index 0000000000..c057047800
--- /dev/null
+++ b/memory/replace/dmd/DMD.h
@@ -0,0 +1,291 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef DMD_h___
+#define DMD_h___
+
+#include <stdarg.h>
+#include <string.h>
+
+#include <utility>
+
+#include "mozilla/DebugOnly.h"
+#include "mozilla/Types.h"
+#include "mozilla/UniquePtr.h"
+#include "replace_malloc_bridge.h"
+
+namespace mozilla {
+
+class JSONWriteFunc;
+
+namespace dmd {
+
+struct Sizes {
+  size_t mStackTracesUsed;
+  size_t mStackTracesUnused;
+  size_t mStackTraceTable;
+  size_t mLiveBlockTable;
+  size_t mDeadBlockTable;
+
+  Sizes() { Clear(); }
+  void Clear() { memset(this, 0, sizeof(Sizes)); }
+};
+
+// See further below for a description of each method. The DMDFuncs class
+// should contain a virtual method for each of them (except IsRunning,
+// which can be inferred from the DMDFuncs singleton existing).
+struct DMDFuncs {
+  virtual void Report(const void*);
+
+  virtual void ReportOnAlloc(const void*);
+
+  virtual void ClearReports();
+
+  virtual void Analyze(UniquePtr<JSONWriteFunc>);
+
+  virtual void SizeOf(Sizes*);
+
+  virtual void StatusMsg(const char*, va_list) MOZ_FORMAT_PRINTF(2, 0);
+
+  virtual void ResetEverything(const char*);
+
+#ifndef REPLACE_MALLOC_IMPL
+  // We deliberately don't use ReplaceMalloc::GetDMDFuncs here, because if we
+  // did, the following would happen.
+  // - The code footprint of each call to Get() larger as GetDMDFuncs ends
+  //   up inlined.
+  // - When no replace-malloc library is loaded, the number of instructions
+  //   executed is equivalent, but don't necessarily fit in the same cache
+  //   line.
+  // - When a non-DMD replace-malloc library is loaded, the overhead is
+  //   higher because there is first a check for the replace malloc bridge
+  //   and then for the DMDFuncs singleton.
+  // Initializing the DMDFuncs singleton on the first access makes the
+  // overhead even worse. Either Get() is inlined and massive, or it isn't
+  // and a simple value check becomes a function call.
+  static DMDFuncs* Get() { return sSingleton.Get(); }
+
+ private:
+  // Wrapper class keeping a pointer to the DMD functions. It is statically
+  // initialized because it needs to be set early enough.
+  // Debug builds also check that it's never accessed before the static
+  // initialization actually occured, which could be the case if some other
+  // static initializer ended up calling into DMD.
+  class Singleton {
+   public:
+    Singleton()
+        : mValue(ReplaceMalloc::GetDMDFuncs())
+#  ifdef DEBUG
+          ,
+          mInitialized(true)
+#  endif
+    {
+    }
+
+    DMDFuncs* Get() {
+      MOZ_ASSERT(mInitialized);
+      return mValue;
+    }
+
+   private:
+    DMDFuncs* mValue;
+#  ifdef DEBUG
+    bool mInitialized;
+#  endif
+  };
+
+  // This singleton pointer must be defined on the program side. In Gecko,
+  // this is done in xpcom/base/nsMemoryInfoDumper.cpp.
+  static /* DMDFuncs:: */ Singleton sSingleton;
+#endif
+};
+
+#ifndef REPLACE_MALLOC_IMPL
+// Mark a heap block as reported by a memory reporter.
+inline void Report(const void* aPtr) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->Report(aPtr);
+  }
+}
+
+// Mark a heap block as reported immediately on allocation.
+inline void ReportOnAlloc(const void* aPtr) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->ReportOnAlloc(aPtr);
+  }
+}
+
+// Clears existing reportedness data from any prior runs of the memory
+// reporters.  The following sequence should be used.
+// - ClearReports()
+// - run the memory reporters
+// - Analyze()
+// This sequence avoids spurious twice-reported warnings.
+inline void ClearReports() {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->ClearReports();
+  }
+}
+
+// Determines which heap blocks have been reported, and dumps JSON output
+// (via |aWriter|) describing the heap.
+//
+// The following sample output contains comments that explain the format and
+// design choices. The output files can be quite large, so a number of
+// decisions were made to minimize size, such as using short property names and
+// omitting properties whenever possible.
+//
+// {
+//   // The version number of the format, which will be incremented each time
+//   // backwards-incompatible changes are made. A mandatory integer.
+//   //
+//   // Version history:
+//   // - 1: Bug 1044709
+//   // - 2: Bug 1094552
+//   // - 3: Bug 1100851
+//   // - 4: Bug 1121830
+//   // - 5: Bug 1253512
+//   "version": 5,
+//
+//   // Information about how DMD was invoked. A mandatory object.
+//   "invocation": {
+//     // The contents of the $DMD environment variable. A string, or |null| if
+//     // $DMD is undefined.
+//     "dmdEnvVar": "--mode=dark-matter",
+//
+//     // The profiling mode. A mandatory string taking one of the following
+//     // values: "live", "dark-matter", "cumulative", "scan".
+//     "mode": "dark-matter",
+//   },
+//
+//   // Details of all analyzed heap blocks. A mandatory array.
+//   "blockList": [
+//     // An example of a heap block.
+//     {
+//       // Requested size, in bytes. This is a mandatory integer.
+//       "req": 3584,
+//
+//       // Requested slop size, in bytes. This is mandatory if it is non-zero,
+//       // but omitted otherwise.
+//       "slop": 512,
+//
+//       // The stack trace at which the block was allocated. An optional
+//       // string that indexes into the "traceTable" object. If omitted, no
+//       // allocation stack trace was recorded for the block.
+//       "alloc": "A",
+//
+//       // One or more stack traces at which this heap block was reported by a
+//       // memory reporter. An optional array that will only be present in
+//       // "dark-matter" mode. The elements are strings that index into
+//       // the "traceTable" object.
+//       "reps": ["B"]
+//
+//       // The number of heap blocks with exactly the above properties. This
+//       // is mandatory if it is greater than one, but omitted otherwise.
+//       // (Blocks with identical properties don't have to be aggregated via
+//       // this property, but it can greatly reduce output file size.)
+//       "num": 5,
+//
+//       // The address of the block. This is mandatory in "scan" mode, but
+//       // omitted otherwise.
+//       "addr": "4e4e4e4e",
+//
+//       // The contents of the block, read one word at a time. This is
+//       // mandatory in "scan" mode for blocks at least one word long, but
+//       // omitted otherwise.
+//       "contents": ["0", "6", "7f7f7f7f", "0"]
+//     }
+//   ],
+//
+//   // The stack traces referenced by elements of the "blockList" array. This
+//   // could be an array, but making it an object makes it easier to see
+//   // which stacks correspond to which references in the "blockList" array.
+//   "traceTable": {
+//     // Each property corresponds to a stack trace mentioned in the "blocks"
+//     // object. Each element is an index into the "frameTable" object.
+//     "A": ["D", "E"],
+//     "B": ["F", "G"]
+//   },
+//
+//   // The stack frames referenced by the "traceTable" object. The
+//   // descriptions can be quite long, so they are stored separately from the
+//   // "traceTable" object so that each one only has to be written once.
+//   // This could also be an array, but again, making it an object makes it
+//   // easier to see which frames correspond to which references in the
+//   // "traceTable" object.
+//   "frameTable": {
+//     // Each property key is a frame key mentioned in the "traceTable" object.
+//     // Each property value is a string containing a frame description. Each
+//     // frame description must be in a format recognized by `fix_stacks.py`,
+//     // which requires a frame number at the start. Because each stack frame
+//     // description in this table can be shared between multiple stack
+//     // traces, we use a dummy value of #00. The proper frame number can be
+//     // reconstructed later by scripts that output stack traces in a
+//     // conventional non-shared format.
+//     "D": "#00: foo (Foo.cpp:123)",
+//     "E": "#00: bar (Bar.cpp:234)",
+//     "F": "#00: baz (Baz.cpp:345)",
+//     "G": "#00: quux (Quux.cpp:456)"
+//   }
+// }
+//
+// Implementation note: normally, this function wouldn't be templated, but in
+// that case, the function is compiled, which makes the destructor for the
+// UniquePtr fire up, and that needs JSONWriteFunc to be fully defined. That,
+// in turn, requires to include JSONWriter.h, which includes
+// double-conversion.h, which ends up breaking various things built with
+// -Werror for various reasons.
+//
+template <typename JSONWriteFunc>
+inline void Analyze(UniquePtr<JSONWriteFunc> aWriteFunc) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->Analyze(std::move(aWriteFunc));
+  }
+}
+
+// Gets the size of various data structures.  Used to implement a memory
+// reporter for DMD.
+inline void SizeOf(Sizes* aSizes) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->SizeOf(aSizes);
+  }
+}
+
+// Prints a status message prefixed with "DMD[<pid>]". Use sparingly.
+MOZ_FORMAT_PRINTF(1, 2)
+inline void StatusMsg(const char* aFmt, ...) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    va_list ap;
+    va_start(ap, aFmt);
+    funcs->StatusMsg(aFmt, ap);
+    va_end(ap);
+  }
+}
+
+// Indicates whether or not DMD is running.
+inline bool IsRunning() { return !!DMDFuncs::Get(); }
+
+// Resets all DMD options and then sets new ones according to those specified
+// in |aOptions|. Also clears all recorded data about allocations. Only used
+// for testing purposes.
+inline void ResetEverything(const char* aOptions) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->ResetEverything(aOptions);
+  }
+}
+#endif
+
+}  // namespace dmd
+}  // namespace mozilla
+
+#endif /* DMD_h___ */
diff --git a/memory/replace/dmd/README b/memory/replace/dmd/README
new file mode 100644
index 0000000000..0b2b9cc1e0
--- /dev/null
+++ b/memory/replace/dmd/README
@@ -0,0 +1,3 @@
+This is DMD.  See
+https://developer.mozilla.org/en-US/docs/Mozilla/Performance/DMD for
+details on how to use it.
diff --git a/memory/replace/dmd/block_analyzer.py b/memory/replace/dmd/block_analyzer.py
new file mode 100644
index 0000000000..dcdfeff7ff
--- /dev/null
+++ b/memory/replace/dmd/block_analyzer.py
@@ -0,0 +1,293 @@
+#!/usr/bin/python
+
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+# From a scan mode DMD log, extract some information about a
+# particular block, such as its allocation stack or which other blocks
+# contain pointers to it. This can be useful when investigating leaks
+# caused by unknown references to refcounted objects.
+
+from __future__ import absolute_import, print_function
+
+import json
+import gzip
+import sys
+import argparse
+import re
+
+
+# The DMD output version this script handles.
+outputVersion = 5
+
+# If --ignore-alloc-fns is specified, stack frames containing functions that
+# match these strings will be removed from the *start* of stack traces. (Once
+# we hit a non-matching frame, any subsequent frames won't be removed even if
+# they do match.)
+allocatorFns = [
+    "malloc (",
+    "replace_malloc",
+    "replace_calloc",
+    "replace_realloc",
+    "replace_memalign",
+    "replace_posix_memalign",
+    "malloc_zone_malloc",
+    "moz_xmalloc",
+    "moz_xcalloc",
+    "moz_xrealloc",
+    "operator new(",
+    "operator new[](",
+    "g_malloc",
+    "g_slice_alloc",
+    "callocCanGC",
+    "reallocCanGC",
+    "vpx_malloc",
+    "vpx_calloc",
+    "vpx_realloc",
+    "vpx_memalign",
+    "js_malloc",
+    "js_calloc",
+    "js_realloc",
+    "pod_malloc",
+    "pod_calloc",
+    "pod_realloc",
+    "nsTArrayInfallibleAllocator::Malloc",
+    "Allocator<ReplaceMallocBase>::malloc(",
+    # This one necessary to fully filter some sequences of allocation functions
+    # that happen in practice. Note that ??? entries that follow non-allocation
+    # functions won't be stripped, as explained above.
+    "???",
+]
+
+####
+
+# Command line arguments
+
+
+def range_1_24(string):
+    value = int(string)
+    if value < 1 or value > 24:
+        msg = "{:s} is not in the range 1..24".format(string)
+        raise argparse.ArgumentTypeError(msg)
+    return value
+
+
+parser = argparse.ArgumentParser(
+    description="Analyze the heap graph to find out things about an object. \
+By default this prints out information about blocks that point to the given block."
+)
+
+parser.add_argument("dmd_log_file_name", help="clamped DMD log file name")
+
+parser.add_argument("block", help="address of the block of interest")
+
+parser.add_argument(
+    "--info",
+    dest="info",
+    action="store_true",
+    default=False,
+    help="Print out information about the block.",
+)
+
+parser.add_argument(
+    "-sfl",
+    "--max-stack-frame-length",
+    type=int,
+    default=150,
+    help="Maximum number of characters to print from each stack frame",
+)
+
+parser.add_argument(
+    "-a",
+    "--ignore-alloc-fns",
+    action="store_true",
+    help="ignore allocation functions at the start of traces",
+)
+
+parser.add_argument(
+    "-f",
+    "--max-frames",
+    type=range_1_24,
+    default=8,
+    help="maximum number of frames to consider in each trace",
+)
+
+parser.add_argument(
+    "-c",
+    "--chain-reports",
+    action="store_true",
+    help="if only one block is found to hold onto the object, report "
+    "the next one, too",
+)
+
+
+####
+
+
+class BlockData:
+    def __init__(self, json_block):
+        self.addr = json_block["addr"]
+
+        if "contents" in json_block:
+            contents = json_block["contents"]
+        else:
+            contents = []
+        self.contents = []
+        for c in contents:
+            self.contents.append(int(c, 16))
+
+        self.req_size = json_block["req"]
+
+        self.alloc_stack = json_block["alloc"]
+
+
+def print_trace_segment(args, stacks, block):
+    (traceTable, frameTable) = stacks
+
+    for l in traceTable[block.alloc_stack]:
+        # The 5: is to remove the bogus leading "#00: " from the stack frame.
+        print("  " + frameTable[l][5 : args.max_stack_frame_length])
+
+
+def show_referrers(args, blocks, stacks, block):
+    visited = set([])
+
+    anyFound = False
+
+    while True:
+        referrers = {}
+
+        for b, data in blocks.iteritems():
+            which_edge = 0
+            for e in data.contents:
+                if e == block:
+                    # 8 is the number of bytes per word on a 64-bit system.
+                    # XXX This means that this output will be wrong for logs from 32-bit systems!
+                    referrers.setdefault(b, []).append(8 * which_edge)
+                    anyFound = True
+                which_edge += 1
+
+        for r in referrers:
+            sys.stdout.write(
+                "0x{} size = {} bytes".format(blocks[r].addr, blocks[r].req_size)
+            )
+            plural = "s" if len(referrers[r]) > 1 else ""
+            sys.stdout.write(
+                " at byte offset"
+                + plural
+                + " "
+                + (", ".join(str(x) for x in referrers[r]))
+            )
+            print
+            print_trace_segment(args, stacks, blocks[r])
+            print
+
+        if args.chain_reports:
+            if len(referrers) == 0:
+                sys.stdout.write("Found no more referrers.\n")
+                break
+            if len(referrers) > 1:
+                sys.stdout.write("Found too many referrers.\n")
+                break
+
+            sys.stdout.write("Chaining to next referrer.\n\n")
+            for r in referrers:
+                block = r
+            if block in visited:
+                sys.stdout.write("Found a loop.\n")
+                break
+            visited.add(block)
+        else:
+            break
+
+    if not anyFound:
+        print("No referrers found.")
+
+
+def show_block_info(args, blocks, stacks, block):
+    b = blocks[block]
+    sys.stdout.write("block: 0x{}\n".format(b.addr))
+    sys.stdout.write("requested size: {} bytes\n".format(b.req_size))
+    sys.stdout.write("\n")
+    sys.stdout.write("block contents: ")
+    for c in b.contents:
+        v = "0" if c == 0 else blocks[c].addr
+        sys.stdout.write("0x{} ".format(v))
+    sys.stdout.write("\n\n")
+    sys.stdout.write("allocation stack:\n")
+    print_trace_segment(args, stacks, b)
+    return
+
+
+def cleanupTraceTable(args, frameTable, traceTable):
+    # Remove allocation functions at the start of traces.
+    if args.ignore_alloc_fns:
+        # Build a regexp that matches every function in allocatorFns.
+        escapedAllocatorFns = map(re.escape, allocatorFns)
+        fn_re = re.compile("|".join(escapedAllocatorFns))
+
+        # Remove allocator fns from each stack trace.
+        for traceKey, frameKeys in traceTable.items():
+            numSkippedFrames = 0
+            for frameKey in frameKeys:
+                frameDesc = frameTable[frameKey]
+                if re.search(fn_re, frameDesc):
+                    numSkippedFrames += 1
+                else:
+                    break
+            if numSkippedFrames > 0:
+                traceTable[traceKey] = frameKeys[numSkippedFrames:]
+
+    # Trim the number of frames.
+    for traceKey, frameKeys in traceTable.items():
+        if len(frameKeys) > args.max_frames:
+            traceTable[traceKey] = frameKeys[: args.max_frames]
+
+
+def loadGraph(options):
+    # Handle gzipped input if necessary.
+    isZipped = options.dmd_log_file_name.endswith(".gz")
+    opener = gzip.open if isZipped else open
+
+    with opener(options.dmd_log_file_name, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    block_list = j["blockList"]
+    blocks = {}
+
+    for json_block in block_list:
+        blocks[int(json_block["addr"], 16)] = BlockData(json_block)
+
+    traceTable = j["traceTable"]
+    frameTable = j["frameTable"]
+
+    cleanupTraceTable(options, frameTable, traceTable)
+
+    return (blocks, (traceTable, frameTable))
+
+
+def analyzeLogs():
+    options = parser.parse_args()
+
+    (blocks, stacks) = loadGraph(options)
+
+    block = int(options.block, 16)
+
+    if block not in blocks:
+        print("Object " + options.block + " not found in traces.")
+        print("It could still be the target of some nodes.")
+        return
+
+    if options.info:
+        show_block_info(options, blocks, stacks, block)
+        return
+
+    show_referrers(options, blocks, stacks, block)
+
+
+if __name__ == "__main__":
+    analyzeLogs()
diff --git a/memory/replace/dmd/dmd.py b/memory/replace/dmd/dmd.py
new file mode 100755
index 0000000000..46e59eefc1
--- /dev/null
+++ b/memory/replace/dmd/dmd.py
@@ -0,0 +1,1002 @@
+#! /usr/bin/env python
+#
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+"""This script analyzes a JSON file emitted by DMD."""
+
+from __future__ import absolute_import, print_function, division
+
+import argparse
+import collections
+import gzip
+import json
+import os
+import platform
+import re
+import shutil
+import sys
+import tempfile
+from bisect import bisect_right
+
+# The DMD output version this script handles.
+outputVersion = 5
+
+# If --ignore-alloc-fns is specified, stack frames containing functions that
+# match these strings will be removed from the *start* of stack traces. (Once
+# we hit a non-matching frame, any subsequent frames won't be removed even if
+# they do match.)
+allocatorFns = [
+    # Matches malloc, replace_malloc, moz_xmalloc, vpx_malloc, js_malloc,
+    # pod_malloc, malloc_zone_*, g_malloc.
+    "malloc",
+    # Matches calloc, replace_calloc, moz_xcalloc, vpx_calloc, js_calloc,
+    # pod_calloc, malloc_zone_calloc, pod_callocCanGC.
+    "calloc",
+    # Matches realloc, replace_realloc, moz_xrealloc, vpx_realloc, js_realloc,
+    # pod_realloc, pod_reallocCanGC.
+    "realloc",
+    # Matches memalign, posix_memalign, replace_memalign, replace_posix_memalign,
+    # moz_xmemalign, vpx_memalign, malloc_zone_memalign.
+    "memalign",
+    "operator new(",
+    "operator new[](",
+    "g_slice_alloc",
+    # This one necessary to fully filter some sequences of allocation functions
+    # that happen in practice. Note that ??? entries that follow non-allocation
+    # functions won't be stripped, as explained above.
+    "???",
+]
+
+
+class Record(object):
+    """A record is an aggregation of heap blocks that have identical stack
+    traces. It can also be used to represent the difference between two
+    records."""
+
+    def __init__(self):
+        self.numBlocks = 0
+        self.reqSize = 0
+        self.slopSize = 0
+        self.usableSize = 0
+        self.allocatedAtDesc = None
+        self.reportedAtDescs = []
+        self.usableSizes = collections.defaultdict(int)
+
+    def isZero(self, args):
+        return (
+            self.numBlocks == 0
+            and self.reqSize == 0
+            and self.slopSize == 0
+            and self.usableSize == 0
+            and len(self.usableSizes) == 0
+        )
+
+    def negate(self):
+        self.numBlocks = -self.numBlocks
+        self.reqSize = -self.reqSize
+        self.slopSize = -self.slopSize
+        self.usableSize = -self.usableSize
+
+        negatedUsableSizes = collections.defaultdict(int)
+        for usableSize, count in self.usableSizes.items():
+            negatedUsableSizes[-usableSize] = count
+        self.usableSizes = negatedUsableSizes
+
+    def subtract(self, r):
+        # We should only be calling this on records with matching stack traces.
+        # Check this.
+        assert self.allocatedAtDesc == r.allocatedAtDesc
+        assert self.reportedAtDescs == r.reportedAtDescs
+
+        self.numBlocks -= r.numBlocks
+        self.reqSize -= r.reqSize
+        self.slopSize -= r.slopSize
+        self.usableSize -= r.usableSize
+
+        usableSizes1 = self.usableSizes
+        usableSizes2 = r.usableSizes
+        usableSizes3 = collections.defaultdict(int)
+        for usableSize in usableSizes1:
+            counts1 = usableSizes1[usableSize]
+            if usableSize in usableSizes2:
+                counts2 = usableSizes2[usableSize]
+                del usableSizes2[usableSize]
+                counts3 = counts1 - counts2
+                if counts3 != 0:
+                    if counts3 < 0:
+                        usableSize = -usableSize
+                        counts3 = -counts3
+                    usableSizes3[usableSize] = counts3
+            else:
+                usableSizes3[usableSize] = counts1
+
+        for usableSize in usableSizes2:
+            usableSizes3[-usableSize] = usableSizes2[usableSize]
+
+        self.usableSizes = usableSizes3
+
+    @staticmethod
+    def cmpByUsableSize(r1, r2):
+        # Sort by usable size, then by req size.
+        return cmp(abs(r1.usableSize), abs(r2.usableSize)) or Record.cmpByReqSize(
+            r1, r2
+        )
+
+    @staticmethod
+    def cmpByReqSize(r1, r2):
+        # Sort by req size.
+        return cmp(abs(r1.reqSize), abs(r2.reqSize))
+
+    @staticmethod
+    def cmpBySlopSize(r1, r2):
+        # Sort by slop size.
+        return cmp(abs(r1.slopSize), abs(r2.slopSize))
+
+    @staticmethod
+    def cmpByNumBlocks(r1, r2):
+        # Sort by block counts, then by usable size.
+        return cmp(abs(r1.numBlocks), abs(r2.numBlocks)) or Record.cmpByUsableSize(
+            r1, r2
+        )
+
+
+sortByChoices = {
+    "usable": Record.cmpByUsableSize,  # the default
+    "req": Record.cmpByReqSize,
+    "slop": Record.cmpBySlopSize,
+    "num-blocks": Record.cmpByNumBlocks,
+}
+
+
+def parseCommandLine():
+    # 24 is the maximum number of frames that DMD will produce.
+    def range_1_24(string):
+        value = int(string)
+        if value < 1 or value > 24:
+            msg = "{:s} is not in the range 1..24".format(string)
+            raise argparse.ArgumentTypeError(msg)
+        return value
+
+    description = """
+Analyze heap data produced by DMD.
+If one file is specified, analyze it; if two files are specified, analyze the
+difference.
+Input files can be gzipped.
+Write to stdout unless -o/--output is specified.
+Stack traces are fixed to show function names, filenames and line numbers
+unless --no-fix-stacks is specified; stack fixing modifies the original file
+and may take some time. If specified, the BREAKPAD_SYMBOLS_PATH environment
+variable is used to find breakpad symbols for stack fixing.
+"""
+    p = argparse.ArgumentParser(description=description)
+
+    p.add_argument(
+        "-o",
+        "--output",
+        type=argparse.FileType("w"),
+        help="output file; stdout if unspecified",
+    )
+
+    p.add_argument(
+        "-f",
+        "--max-frames",
+        type=range_1_24,
+        default=8,
+        help="maximum number of frames to consider in each trace",
+    )
+
+    p.add_argument(
+        "-s",
+        "--sort-by",
+        choices=sortByChoices.keys(),
+        default="usable",
+        help="sort the records by a particular metric",
+    )
+
+    p.add_argument(
+        "-a",
+        "--ignore-alloc-fns",
+        action="store_true",
+        help="ignore allocation functions at the start of traces",
+    )
+
+    p.add_argument("--no-fix-stacks", action="store_true", help="do not fix stacks")
+
+    p.add_argument(
+        "--clamp-contents",
+        action="store_true",
+        help="for a scan mode log, clamp addresses to the start of live blocks, "
+        "or zero if not in one",
+    )
+
+    p.add_argument(
+        "--print-clamp-stats",
+        action="store_true",
+        help="print information about the results of pointer clamping; mostly "
+        "useful for debugging clamping",
+    )
+
+    p.add_argument(
+        "--filter-stacks-for-testing",
+        action="store_true",
+        help="filter stack traces; only useful for testing purposes",
+    )
+
+    p.add_argument(
+        "--allocation-filter",
+        help="Only print entries that have a stack that matches the filter",
+    )
+
+    p.add_argument("input_file", help="a file produced by DMD")
+
+    p.add_argument(
+        "input_file2",
+        nargs="?",
+        help="a file produced by DMD; if present, it is diff'd with input_file",
+    )
+
+    return p.parse_args(sys.argv[1:])
+
+
+# Fix stacks if necessary: first write the output to a tempfile, then replace
+# the original file with it.
+def fixStackTraces(inputFilename, isZipped, opener):
+    # This append() call is needed to make the import statements work when this
+    # script is installed as a symlink.
+    sys.path.append(os.path.dirname(__file__))
+
+    bpsyms = os.environ.get("BREAKPAD_SYMBOLS_PATH", None)
+    sysname = platform.system()
+    if bpsyms and os.path.exists(bpsyms):
+        import fix_stacks as fixModule
+
+        def fix(line):
+            return fixModule.fixSymbols(line, jsonMode=True, breakpadSymsDir=bpsyms)
+
+    elif sysname in ("Linux", "Darwin", "Windows"):
+        import fix_stacks as fixModule
+
+        def fix(line):
+            return fixModule.fixSymbols(line, jsonMode=True)
+
+    else:
+        return
+
+    # Fix stacks, writing output to a temporary file, and then overwrite the
+    # original file.
+    tmpFile = tempfile.NamedTemporaryFile(delete=False)
+
+    # If the input is gzipped, then the output (written initially to |tmpFile|)
+    # should be gzipped as well.
+    #
+    # And we want to set its pre-gzipped filename to '' rather than the name of
+    # the temporary file, so that programs like the Unix 'file' utility don't
+    # say that it was called 'tmp6ozTxE' (or something like that) before it was
+    # zipped. So that explains the |filename=''| parameter.
+    #
+    # But setting the filename like that clobbers |tmpFile.name|, so we must
+    # get that now in order to move |tmpFile| at the end.
+    tmpFilename = tmpFile.name
+    if isZipped:
+        tmpFile = gzip.GzipFile(filename="", fileobj=tmpFile)
+
+    with opener(inputFilename, "rb") as inputFile:
+        for line in inputFile:
+            tmpFile.write(fix(line))
+
+    tmpFile.close()
+
+    shutil.move(tmpFilename, inputFilename)
+
+
+def getDigestFromFile(args, inputFile):
+    # Handle gzipped input if necessary.
+    isZipped = inputFile.endswith(".gz")
+    opener = gzip.open if isZipped else open
+
+    # Fix stack traces unless otherwise instructed.
+    if not args.no_fix_stacks:
+        fixStackTraces(inputFile, isZipped, opener)
+
+    if args.clamp_contents:
+        clampBlockList(args, inputFile, isZipped, opener)
+
+    with opener(inputFile, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    # Extract the main parts of the JSON object.
+    invocation = j["invocation"]
+    dmdEnvVar = invocation["dmdEnvVar"]
+    mode = invocation["mode"]
+    blockList = j["blockList"]
+    traceTable = j["traceTable"]
+    frameTable = j["frameTable"]
+
+    # Insert the necessary entries for unrecorded stack traces. Note that 'ut'
+    # and 'uf' will not overlap with any keys produced by DMD's
+    # ToIdStringConverter::Base32() function.
+    unrecordedTraceID = "ut"
+    unrecordedFrameID = "uf"
+    traceTable[unrecordedTraceID] = [unrecordedFrameID]
+    frameTable[
+        unrecordedFrameID
+    ] = "#00: (no stack trace recorded due to --stacks=partial)"
+
+    # For the purposes of this script, 'scan' behaves like 'live'.
+    if mode == "scan":
+        mode = "live"
+
+    if mode not in ["live", "dark-matter", "cumulative"]:
+        raise Exception("bad 'mode' property: '{:s}'".format(mode))
+
+    # Remove allocation functions at the start of traces.
+    if args.ignore_alloc_fns:
+        # Build a regexp that matches every function in allocatorFns.
+        escapedAllocatorFns = map(re.escape, allocatorFns)
+        fn_re = re.compile("|".join(escapedAllocatorFns))
+
+        # Remove allocator fns from each stack trace.
+        for traceKey, frameKeys in traceTable.items():
+            numSkippedFrames = 0
+            for frameKey in frameKeys:
+                frameDesc = frameTable[frameKey]
+                if re.search(fn_re, frameDesc):
+                    numSkippedFrames += 1
+                else:
+                    break
+            if numSkippedFrames > 0:
+                traceTable[traceKey] = frameKeys[numSkippedFrames:]
+
+    # Trim the number of frames.
+    for traceKey, frameKeys in traceTable.items():
+        if len(frameKeys) > args.max_frames:
+            traceTable[traceKey] = frameKeys[: args.max_frames]
+
+    def buildTraceDescription(traceTable, frameTable, traceKey):
+        frameKeys = traceTable[traceKey]
+        fmt = "    #{:02d}{:}"
+
+        if args.filter_stacks_for_testing:
+            # This option is used by `test_dmd.js`, which runs the code in
+            # `SmokeDMD.cpp`. When running that test, there is too much
+            # variation in the stack traces across different machines and
+            # platforms to do exact output matching. However, every stack trace
+            # should have at least three frames that contain `DMD` (in one of
+            # `DMD.cpp`, `SmokeDMD.cpp`, `SmokeDMD`, or `SmokeDMD.exe`). Some
+            # example frames from automation (where `..` indicates excised path
+            # segments):
+            #
+            # Linux debug, with stack fixing using breakpad syms:
+            # `#01: replace_realloc(void*, unsigned long) [../dmd/DMD.cpp:1110]`
+            #
+            # Linux opt, with native stack fixing:
+            # `#02: TestFull(char const*, int, char const*, int) (../dmd/test/SmokeDMD.cpp:165)`
+            #
+            # Mac opt, with native stack fixing:
+            # `#03: RunTests() (../build/tests/bin/SmokeDMD + 0x21f9)`
+            #
+            # Windows opt, with native stack fixing failing due to a missing PDB:
+            # `#04: ??? (..\\build\\tests\\bin\\SmokeDMD.exe + 0x1c58)`
+            #
+            # If we see three such frames, we replace the entire stack trace
+            # with a single, predictable frame. This imprecise matching will at
+            # least detect if stack fixing fails completely.
+            dmd_frame_matches = 0
+            for frameKey in frameKeys:
+                frameDesc = frameTable[frameKey]
+                if "DMD" in frameDesc:
+                    dmd_frame_matches += 1
+                    if dmd_frame_matches >= 3:
+                        return [fmt.format(1, ": ... DMD.cpp ...")]
+
+        # The frame number is always '#00' (see DMD.h for why), so we have to
+        # replace that with the correct frame number.
+        desc = []
+        for n, frameKey in enumerate(traceTable[traceKey], start=1):
+            desc.append(fmt.format(n, frameTable[frameKey][3:]))
+        return desc
+
+    # Aggregate blocks into records. All sufficiently similar blocks go into a
+    # single record.
+
+    if mode in ["live", "cumulative"]:
+        liveOrCumulativeRecords = collections.defaultdict(Record)
+    elif mode == "dark-matter":
+        unreportedRecords = collections.defaultdict(Record)
+        onceReportedRecords = collections.defaultdict(Record)
+        twiceReportedRecords = collections.defaultdict(Record)
+
+    heapUsableSize = 0
+    heapBlocks = 0
+
+    recordKeyPartCache = {}
+
+    for block in blockList:
+        # For each block we compute a |recordKey|, and all blocks with the same
+        # |recordKey| are aggregated into a single record. The |recordKey| is
+        # derived from the block's 'alloc' and 'reps' (if present) stack
+        # traces.
+        #
+        # We use frame descriptions (e.g. "#00: foo (X.cpp:99)") when comparing
+        # traces for equality. We can't use trace keys or frame keys because
+        # they're not comparable across different DMD runs (which is relevant
+        # when doing diffs).
+        #
+        # Using frame descriptions also fits in with the stack trimming done
+        # for --max-frames, which requires that stack traces with common
+        # beginnings but different endings to be considered equivalent. E.g. if
+        # we have distinct traces T1:[A:D1,B:D2,C:D3] and T2:[X:D1,Y:D2,Z:D4]
+        # and we trim the final frame of each they should be considered
+        # equivalent because the untrimmed frame descriptions (D1 and D2)
+        # match.
+        #
+        # Having said all that, during a single invocation of dmd.py on a
+        # single DMD file, for a single frameKey value the record key will
+        # always be the same, and we might encounter it 1000s of times. So we
+        # cache prior results for speed.
+        def makeRecordKeyPart(traceKey):
+            if traceKey in recordKeyPartCache:
+                return recordKeyPartCache[traceKey]
+
+            recordKeyPart = str(
+                map(lambda frameKey: frameTable[frameKey], traceTable[traceKey])
+            )
+            recordKeyPartCache[traceKey] = recordKeyPart
+            return recordKeyPart
+
+        allocatedAtTraceKey = block.get("alloc", unrecordedTraceID)
+        if mode in ["live", "cumulative"]:
+            recordKey = makeRecordKeyPart(allocatedAtTraceKey)
+            records = liveOrCumulativeRecords
+        elif mode == "dark-matter":
+            recordKey = makeRecordKeyPart(allocatedAtTraceKey)
+            if "reps" in block:
+                reportedAtTraceKeys = block["reps"]
+                for reportedAtTraceKey in reportedAtTraceKeys:
+                    recordKey += makeRecordKeyPart(reportedAtTraceKey)
+                if len(reportedAtTraceKeys) == 1:
+                    records = onceReportedRecords
+                else:
+                    records = twiceReportedRecords
+            else:
+                records = unreportedRecords
+
+        record = records[recordKey]
+
+        if "req" not in block:
+            raise Exception("'req' property missing in block'")
+
+        reqSize = block["req"]
+        slopSize = block.get("slop", 0)
+
+        if "num" in block:
+            num = block["num"]
+        else:
+            num = 1
+
+        usableSize = reqSize + slopSize
+        heapUsableSize += num * usableSize
+        heapBlocks += num
+
+        record.numBlocks += num
+        record.reqSize += num * reqSize
+        record.slopSize += num * slopSize
+        record.usableSize += num * usableSize
+        if record.allocatedAtDesc is None:
+            record.allocatedAtDesc = buildTraceDescription(
+                traceTable, frameTable, allocatedAtTraceKey
+            )
+
+        if mode in ["live", "cumulative"]:
+            pass
+        elif mode == "dark-matter":
+            if "reps" in block and record.reportedAtDescs == []:
+
+                def f(k):
+                    return buildTraceDescription(traceTable, frameTable, k)
+
+                record.reportedAtDescs = map(f, reportedAtTraceKeys)
+        record.usableSizes[usableSize] += num
+
+    # All the processed data for a single DMD file is called a "digest".
+    digest = {}
+    digest["dmdEnvVar"] = dmdEnvVar
+    digest["mode"] = mode
+    digest["heapUsableSize"] = heapUsableSize
+    digest["heapBlocks"] = heapBlocks
+    if mode in ["live", "cumulative"]:
+        digest["liveOrCumulativeRecords"] = liveOrCumulativeRecords
+    elif mode == "dark-matter":
+        digest["unreportedRecords"] = unreportedRecords
+        digest["onceReportedRecords"] = onceReportedRecords
+        digest["twiceReportedRecords"] = twiceReportedRecords
+    return digest
+
+
+def diffRecords(args, records1, records2):
+    records3 = {}
+
+    # Process records1.
+    for k in records1:
+        r1 = records1[k]
+        if k in records2:
+            # This record is present in both records1 and records2.
+            r2 = records2[k]
+            del records2[k]
+            r2.subtract(r1)
+            if not r2.isZero(args):
+                records3[k] = r2
+        else:
+            # This record is present only in records1.
+            r1.negate()
+            records3[k] = r1
+
+    for k in records2:
+        # This record is present only in records2.
+        records3[k] = records2[k]
+
+    return records3
+
+
+def diffDigests(args, d1, d2):
+    if d1["mode"] != d2["mode"]:
+        raise Exception("the input files have different 'mode' properties")
+
+    d3 = {}
+    d3["dmdEnvVar"] = (d1["dmdEnvVar"], d2["dmdEnvVar"])
+    d3["mode"] = d1["mode"]
+    d3["heapUsableSize"] = d2["heapUsableSize"] - d1["heapUsableSize"]
+    d3["heapBlocks"] = d2["heapBlocks"] - d1["heapBlocks"]
+    if d1["mode"] in ["live", "cumulative"]:
+        d3["liveOrCumulativeRecords"] = diffRecords(
+            args, d1["liveOrCumulativeRecords"], d2["liveOrCumulativeRecords"]
+        )
+    elif d1["mode"] == "dark-matter":
+        d3["unreportedRecords"] = diffRecords(
+            args, d1["unreportedRecords"], d2["unreportedRecords"]
+        )
+        d3["onceReportedRecords"] = diffRecords(
+            args, d1["onceReportedRecords"], d2["onceReportedRecords"]
+        )
+        d3["twiceReportedRecords"] = diffRecords(
+            args, d1["twiceReportedRecords"], d2["twiceReportedRecords"]
+        )
+    return d3
+
+
+def printDigest(args, digest):
+    dmdEnvVar = digest["dmdEnvVar"]
+    mode = digest["mode"]
+    heapUsableSize = digest["heapUsableSize"]
+    heapBlocks = digest["heapBlocks"]
+    if mode in ["live", "cumulative"]:
+        liveOrCumulativeRecords = digest["liveOrCumulativeRecords"]
+    elif mode == "dark-matter":
+        unreportedRecords = digest["unreportedRecords"]
+        onceReportedRecords = digest["onceReportedRecords"]
+        twiceReportedRecords = digest["twiceReportedRecords"]
+
+    separator = "#" + "-" * 65 + "\n"
+
+    def number(n):
+        """Format a number with comma as a separator."""
+        return "{:,d}".format(n)
+
+    def perc(m, n):
+        return 0 if n == 0 else (100 * m / n)
+
+    def plural(n):
+        return "" if n == 1 else "s"
+
+    # Prints to stdout, or to file if -o/--output was specified.
+    def out(*arguments, **kwargs):
+        print(*arguments, file=args.output, **kwargs)
+
+    def printStack(traceDesc):
+        for frameDesc in traceDesc:
+            out(frameDesc)
+
+    def printRecords(recordKind, records, heapUsableSize):
+        RecordKind = recordKind.capitalize()
+        out(separator)
+        numRecords = len(records)
+        cmpRecords = sortByChoices[args.sort_by]
+        sortedRecords = sorted(records.values(), cmp=cmpRecords, reverse=True)
+        kindBlocks = 0
+        kindUsableSize = 0
+        maxRecord = 1000
+
+        if args.allocation_filter:
+            sortedRecords = list(
+                filter(
+                    lambda x: any(
+                        map(lambda y: args.allocation_filter in y, x.allocatedAtDesc)
+                    ),
+                    sortedRecords,
+                )
+            )
+
+        # First iteration: get totals, etc.
+        for record in sortedRecords:
+            kindBlocks += record.numBlocks
+            kindUsableSize += record.usableSize
+
+        # Second iteration: print.
+        if numRecords == 0:
+            out("# no {:} heap blocks\n".format(recordKind))
+
+        kindCumulativeUsableSize = 0
+        for i, record in enumerate(sortedRecords, start=1):
+            # Stop printing at the |maxRecord|th record.
+            if i == maxRecord:
+                out(
+                    "# {:}: stopping after {:,d} heap block records\n".format(
+                        RecordKind, i
+                    )
+                )
+                break
+
+            kindCumulativeUsableSize += record.usableSize
+
+            out(RecordKind + " {")
+            out(
+                "  {:} block{:} in heap block record {:,d} of {:,d}".format(
+                    number(record.numBlocks), plural(record.numBlocks), i, numRecords
+                )
+            )
+            out(
+                "  {:} bytes ({:} requested / {:} slop)".format(
+                    number(record.usableSize),
+                    number(record.reqSize),
+                    number(record.slopSize),
+                )
+            )
+
+            usableSizes = sorted(
+                record.usableSizes.items(), key=lambda x: abs(x[0]), reverse=True
+            )
+            hasSingleBlock = len(usableSizes) == 1 and usableSizes[0][1] == 1
+
+            if not hasSingleBlock:
+                out("  Individual block sizes: ", end="")
+                if len(usableSizes) == 0:
+                    out("(no change)", end="")
+                else:
+                    isFirst = True
+                    for usableSize, count in usableSizes:
+                        if not isFirst:
+                            out("; ", end="")
+                        out("{:}".format(number(usableSize)), end="")
+                        if count > 1:
+                            out(" x {:,d}".format(count), end="")
+                        isFirst = False
+                out()
+
+            out(
+                "  {:4.2f}% of the heap ({:4.2f}% cumulative)".format(
+                    perc(record.usableSize, heapUsableSize),
+                    perc(kindCumulativeUsableSize, heapUsableSize),
+                )
+            )
+            if mode in ["live", "cumulative"]:
+                pass
+            elif mode == "dark-matter":
+                out(
+                    "  {:4.2f}% of {:} ({:4.2f}% cumulative)".format(
+                        perc(record.usableSize, kindUsableSize),
+                        recordKind,
+                        perc(kindCumulativeUsableSize, kindUsableSize),
+                    )
+                )
+            out("  Allocated at {")
+            printStack(record.allocatedAtDesc)
+            out("  }")
+            if mode in ["live", "cumulative"]:
+                pass
+            elif mode == "dark-matter":
+                for n, reportedAtDesc in enumerate(record.reportedAtDescs):
+                    again = "again " if n > 0 else ""
+                    out("  Reported {:}at {{".format(again))
+                    printStack(reportedAtDesc)
+                    out("  }")
+            out("}\n")
+
+        return (kindUsableSize, kindBlocks)
+
+    def printInvocation(n, dmdEnvVar, mode):
+        out("Invocation{:} {{".format(n))
+        if dmdEnvVar is None:
+            out("  $DMD is undefined")
+        else:
+            out("  $DMD = '" + dmdEnvVar + "'")
+        out("  Mode = '" + mode + "'")
+        out("}\n")
+
+    # Print command line. Strip dirs so the output is deterministic, which is
+    # needed for testing.
+    out(separator, end="")
+    out("# " + " ".join(map(os.path.basename, sys.argv)) + "\n")
+
+    # Print invocation(s).
+    if type(dmdEnvVar) is not tuple:
+        printInvocation("", dmdEnvVar, mode)
+    else:
+        printInvocation(" 1", dmdEnvVar[0], mode)
+        printInvocation(" 2", dmdEnvVar[1], mode)
+
+    # Print records.
+    if mode in ["live", "cumulative"]:
+        liveOrCumulativeUsableSize, liveOrCumulativeBlocks = printRecords(
+            mode, liveOrCumulativeRecords, heapUsableSize
+        )
+    elif mode == "dark-matter":
+        twiceReportedUsableSize, twiceReportedBlocks = printRecords(
+            "twice-reported", twiceReportedRecords, heapUsableSize
+        )
+
+        unreportedUsableSize, unreportedBlocks = printRecords(
+            "unreported", unreportedRecords, heapUsableSize
+        )
+
+        onceReportedUsableSize, onceReportedBlocks = printRecords(
+            "once-reported", onceReportedRecords, heapUsableSize
+        )
+
+    # Print summary.
+    out(separator)
+    out("Summary {")
+    if mode in ["live", "cumulative"]:
+        out(
+            "  Total: {:} bytes in {:} blocks".format(
+                number(liveOrCumulativeUsableSize), number(liveOrCumulativeBlocks)
+            )
+        )
+    elif mode == "dark-matter":
+        fmt = "  {:15} {:>12} bytes ({:6.2f}%) in {:>7} blocks ({:6.2f}%)"
+        out(fmt.format("Total:", number(heapUsableSize), 100, number(heapBlocks), 100))
+        out(
+            fmt.format(
+                "Unreported:",
+                number(unreportedUsableSize),
+                perc(unreportedUsableSize, heapUsableSize),
+                number(unreportedBlocks),
+                perc(unreportedBlocks, heapBlocks),
+            )
+        )
+        out(
+            fmt.format(
+                "Once-reported:",
+                number(onceReportedUsableSize),
+                perc(onceReportedUsableSize, heapUsableSize),
+                number(onceReportedBlocks),
+                perc(onceReportedBlocks, heapBlocks),
+            )
+        )
+        out(
+            fmt.format(
+                "Twice-reported:",
+                number(twiceReportedUsableSize),
+                perc(twiceReportedUsableSize, heapUsableSize),
+                number(twiceReportedBlocks),
+                perc(twiceReportedBlocks, heapBlocks),
+            )
+        )
+    out("}\n")
+
+
+#############################
+# Pretty printer for DMD JSON
+#############################
+
+
+def prettyPrintDmdJson(out, j):
+    out.write("{\n")
+
+    out.write(' "version": {0},\n'.format(j["version"]))
+    out.write(' "invocation": ')
+    json.dump(j["invocation"], out, sort_keys=True)
+    out.write(",\n")
+
+    out.write(' "blockList": [')
+    first = True
+    for b in j["blockList"]:
+        out.write("" if first else ",")
+        out.write("\n  ")
+        json.dump(b, out, sort_keys=True)
+        first = False
+    out.write("\n ],\n")
+
+    out.write(' "traceTable": {')
+    first = True
+    for k, l in j["traceTable"].iteritems():
+        out.write("" if first else ",")
+        out.write('\n  "{0}": {1}'.format(k, json.dumps(l)))
+        first = False
+    out.write("\n },\n")
+
+    out.write(' "frameTable": {')
+    first = True
+    for k, v in j["frameTable"].iteritems():
+        out.write("" if first else ",")
+        out.write('\n  "{0}": {1}'.format(k, json.dumps(v)))
+        first = False
+    out.write("\n }\n")
+
+    out.write("}\n")
+
+
+##################################################################
+# Code for clamping addresses using conservative pointer analysis.
+##################################################################
+
+# Start is the address of the first byte of the block, while end is
+# the address of the first byte after the final byte in the block.
+class AddrRange:
+    def __init__(self, block, length):
+        self.block = block
+        self.start = int(block, 16)
+        self.length = length
+        self.end = self.start + self.length
+
+        assert self.start > 0
+        assert length >= 0
+
+
+class ClampStats:
+    def __init__(self):
+        # Number of pointers already pointing to the start of a block.
+        self.startBlockPtr = 0
+
+        # Number of pointers pointing to the middle of a block. These
+        # are clamped to the start of the block they point into.
+        self.midBlockPtr = 0
+
+        # Number of null pointers.
+        self.nullPtr = 0
+
+        # Number of non-null pointers that didn't point into the middle
+        # of any blocks. These are clamped to null.
+        self.nonNullNonBlockPtr = 0
+
+    def clampedBlockAddr(self, sameAddress):
+        if sameAddress:
+            self.startBlockPtr += 1
+        else:
+            self.midBlockPtr += 1
+
+    def nullAddr(self):
+        self.nullPtr += 1
+
+    def clampedNonBlockAddr(self):
+        self.nonNullNonBlockPtr += 1
+
+    def log(self):
+        sys.stderr.write("Results:\n")
+        sys.stderr.write(
+            "  Number of pointers already pointing to start of blocks: "
+            + str(self.startBlockPtr)
+            + "\n"
+        )
+        sys.stderr.write(
+            "  Number of pointers clamped to start of blocks: "
+            + str(self.midBlockPtr)
+            + "\n"
+        )
+        sys.stderr.write(
+            "  Number of non-null pointers not pointing into blocks "
+            "clamped to null: " + str(self.nonNullNonBlockPtr) + "\n"
+        )
+        sys.stderr.write("  Number of null pointers: " + str(self.nullPtr) + "\n")
+
+
+# Search the block ranges array for a block that address points into.
+# The search is carried out in an array of starting addresses for each blocks
+# because it is faster.
+def clampAddress(blockRanges, blockStarts, clampStats, address):
+    i = bisect_right(blockStarts, address)
+
+    # Any addresses completely out of the range should have been eliminated already.
+    assert i > 0
+    r = blockRanges[i - 1]
+    assert r.start <= address
+
+    if address >= r.end:
+        assert address < blockRanges[i].start
+        clampStats.clampedNonBlockAddr()
+        return "0"
+
+    clampStats.clampedBlockAddr(r.start == address)
+    return r.block
+
+
+def clampBlockList(args, inputFileName, isZipped, opener):
+    # XXX This isn't very efficient because we end up reading and writing
+    # the file multiple times.
+    with opener(inputFileName, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    # Check that the invocation is reasonable for contents clamping.
+    invocation = j["invocation"]
+    if invocation["mode"] != "scan":
+        raise Exception("Log was taken in mode " + invocation["mode"] + " not scan")
+
+    sys.stderr.write("Creating block range list.\n")
+    blockList = j["blockList"]
+    blockRanges = []
+    for block in blockList:
+        blockRanges.append(AddrRange(block["addr"], block["req"]))
+    blockRanges.sort(key=lambda r: r.start)
+
+    # Make sure there are no overlapping blocks.
+    prevRange = blockRanges[0]
+    for currRange in blockRanges[1:]:
+        assert prevRange.end <= currRange.start
+        prevRange = currRange
+
+    sys.stderr.write("Clamping block contents.\n")
+    clampStats = ClampStats()
+    firstAddr = blockRanges[0].start
+    lastAddr = blockRanges[-1].end
+
+    blockStarts = []
+    for r in blockRanges:
+        blockStarts.append(r.start)
+
+    for block in blockList:
+        # Small blocks don't have any contents.
+        if "contents" not in block:
+            continue
+
+        cont = block["contents"]
+        for i in range(len(cont)):
+            address = int(cont[i], 16)
+
+            if address == 0:
+                clampStats.nullAddr()
+                continue
+
+            # If the address is before the first block or after the last
+            # block then it can't be within a block.
+            if address < firstAddr or address >= lastAddr:
+                clampStats.clampedNonBlockAddr()
+                cont[i] = "0"
+                continue
+
+            cont[i] = clampAddress(blockRanges, blockStarts, clampStats, address)
+
+        # Remove any trailing nulls.
+        while len(cont) and cont[-1] == "0":
+            cont.pop()
+
+    if args.print_clamp_stats:
+        clampStats.log()
+
+    sys.stderr.write("Saving file.\n")
+    tmpFile = tempfile.NamedTemporaryFile(delete=False)
+    tmpFilename = tmpFile.name
+    if isZipped:
+        tmpFile = gzip.GzipFile(filename="", fileobj=tmpFile)
+    prettyPrintDmdJson(tmpFile, j)
+    tmpFile.close()
+    shutil.move(tmpFilename, inputFileName)
+
+
+def main():
+    args = parseCommandLine()
+    digest = getDigestFromFile(args, args.input_file)
+    if args.input_file2:
+        digest2 = getDigestFromFile(args, args.input_file2)
+        digest = diffDigests(args, digest, digest2)
+    printDigest(args, digest)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/memory/replace/dmd/moz.build b/memory/replace/dmd/moz.build
new file mode 100644
index 0000000000..6f3121df48
--- /dev/null
+++ b/memory/replace/dmd/moz.build
@@ -0,0 +1,37 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+EXPORTS += [
+    "DMD.h",
+]
+
+UNIFIED_SOURCES += [
+    "DMD.cpp",
+]
+
+if not CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    UNIFIED_SOURCES += [
+        "/mfbt/HashFunctions.cpp",
+        "/mfbt/JSONWriter.cpp",
+        "/mfbt/Poison.cpp",
+        "/mozglue/misc/StackWalk.cpp",
+    ]
+    if CONFIG["OS_ARCH"] == "WINNT":
+        OS_LIBS += [
+            "dbghelp",
+        ]
+
+ReplaceMalloc("dmd")
+
+DEFINES["MOZ_NO_MOZALLOC"] = True
+DEFINES["IMPL_MFBT"] = True
+
+if CONFIG["MOZ_OPTIMIZE"]:
+    DEFINES["MOZ_OPTIMIZE"] = True
+
+DisableStlWrapping()
+
+TEST_DIRS += ["test"]
diff --git a/memory/replace/dmd/test/.eslintrc.js b/memory/replace/dmd/test/.eslintrc.js
new file mode 100644
index 0000000000..69e89d0054
--- /dev/null
+++ b/memory/replace/dmd/test/.eslintrc.js
@@ -0,0 +1,5 @@
+"use strict";
+
+module.exports = {
+  extends: ["plugin:mozilla/xpcshell-test"],
+};
diff --git a/memory/replace/dmd/test/SmokeDMD.cpp b/memory/replace/dmd/test/SmokeDMD.cpp
new file mode 100644
index 0000000000..e36ae57d50
--- /dev/null
+++ b/memory/replace/dmd/test/SmokeDMD.cpp
@@ -0,0 +1,376 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this file,
+ * You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// This program is used by the DMD xpcshell test. It is run under DMD and
+// produces some output. The xpcshell test then post-processes and checks this
+// output.
+//
+// Note that this file does not have "Test" or "test" in its name, because that
+// will cause the build system to not record breakpad symbols for it, which
+// will stop the post-processing (which includes stack fixing) from working
+// correctly.
+
+// This is required on some systems such as Fedora to allow
+// building with -O0 together with --warnings-as-errors due to
+// a check in /usr/include/features.h
+#undef _FORTIFY_SOURCE
+
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/JSONWriter.h"
+#include "mozilla/UniquePtr.h"
+#include "DMD.h"
+
+using mozilla::MakeUnique;
+using namespace mozilla::dmd;
+
+DMDFuncs::Singleton DMDFuncs::sSingleton;
+
+class FpWriteFunc : public mozilla::JSONWriteFunc {
+ public:
+  explicit FpWriteFunc(const char* aFilename) {
+    mFp = fopen(aFilename, "w");
+    if (!mFp) {
+      fprintf(stderr, "SmokeDMD: can't create %s file: %s\n", aFilename,
+              strerror(errno));
+      exit(1);
+    }
+  }
+
+  ~FpWriteFunc() { fclose(mFp); }
+
+  void Write(const mozilla::Span<const char>& aStr) override {
+    for (const char c : aStr) {
+      fputc(c, mFp);
+    }
+  }
+
+ private:
+  FILE* mFp;
+};
+
+// This stops otherwise-unused variables from being optimized away.
+static void UseItOrLoseIt(void* aPtr, int aSeven) {
+  char buf[64];
+  int n = sprintf(buf, "%p\n", aPtr);
+  if (n == 20 + aSeven) {
+    fprintf(stderr, "well, that is surprising");
+  }
+}
+
+// This function checks that heap blocks that have the same stack trace but
+// different (or no) reporters get aggregated separately.
+void Foo(int aSeven) {
+  char* a[6];
+  for (int i = 0; i < aSeven - 1; i++) {
+    a[i] = (char*)malloc(128 - 16 * i);
+    UseItOrLoseIt(a[i], aSeven);
+  }
+
+  // Oddly, some versions of clang will cause identical stack traces to be
+  // generated for adjacent calls to Report(), which breaks the test. Inserting
+  // the UseItOrLoseIt() calls in between is enough to prevent this.
+
+  Report(a[2]);  // reported
+
+  UseItOrLoseIt(a[2], aSeven);
+
+  for (int i = 0; i < aSeven - 5; i++) {
+    Report(a[i]);  // reported
+    UseItOrLoseIt(a[i], aSeven);
+  }
+
+  UseItOrLoseIt(a[2], aSeven);
+
+  Report(a[3]);  // reported
+
+  // a[4], a[5] unreported
+}
+
+void TestEmpty(const char* aTestName, const char* aMode) {
+  char filename[128];
+  sprintf(filename, "complete-%s-%s.json", aTestName, aMode);
+  auto f = MakeUnique<FpWriteFunc>(filename);
+
+  char options[128];
+  sprintf(options, "--mode=%s --stacks=full", aMode);
+  ResetEverything(options);
+
+  // Zero for everything.
+  Analyze(std::move(f));
+}
+
+void TestFull(const char* aTestName, int aNum, const char* aMode, int aSeven) {
+  char filename[128];
+  sprintf(filename, "complete-%s%d-%s.json", aTestName, aNum, aMode);
+  auto f = MakeUnique<FpWriteFunc>(filename);
+
+  // The --show-dump-stats=yes is there just to give that option some basic
+  // testing, e.g. ensure it doesn't crash. It's hard to test much beyond that.
+  char options[128];
+  sprintf(options, "--mode=%s --stacks=full --show-dump-stats=yes", aMode);
+  ResetEverything(options);
+
+  // Analyze 1: 1 freed, 9 out of 10 unreported.
+  // Analyze 2: still present and unreported.
+  int i;
+  char* a = nullptr;
+  for (i = 0; i < aSeven + 3; i++) {
+    a = (char*)malloc(100);
+    UseItOrLoseIt(a, aSeven);
+  }
+  free(a);
+
+  // A no-op.
+  free(nullptr);
+
+  // Note: 16 bytes is the smallest requested size that gives consistent
+  // behaviour across all platforms with jemalloc.
+  // Analyze 1: reported.
+  // Analyze 2: thrice-reported.
+  char* a2 = (char*)malloc(16);
+  Report(a2);
+
+  // Analyze 1: reported.
+  // Analyze 2: reportedness carries over, due to ReportOnAlloc.
+  char* b = (char*)malloc(10);
+  ReportOnAlloc(b);
+
+  // ReportOnAlloc, then freed.
+  // Analyze 1: freed, irrelevant.
+  // Analyze 2: freed, irrelevant.
+  char* b2 = (char*)malloc(16);
+  ReportOnAlloc(b2);
+  free(b2);
+
+  // Analyze 1: reported 4 times.
+  // Analyze 2: freed, irrelevant.
+  char* c = (char*)calloc(10, 3);
+  Report(c);
+  for (int i = 0; i < aSeven - 4; i++) {
+    Report(c);
+  }
+
+  // Analyze 1: ignored.
+  // Analyze 2: irrelevant.
+  Report((void*)(intptr_t)i);
+
+  // jemalloc rounds this up to 8192.
+  // Analyze 1: reported.
+  // Analyze 2: freed.
+  char* e = (char*)malloc(4096);
+  e = (char*)realloc(e, 7169);
+  Report(e);
+
+  // First realloc is like malloc;  second realloc is shrinking.
+  // Analyze 1: reported.
+  // Analyze 2: re-reported.
+  char* e2 = (char*)realloc(nullptr, 1024);
+  e2 = (char*)realloc(e2, 512);
+  Report(e2);
+
+  // First realloc is like malloc;  second realloc creates a min-sized block.
+  // XXX: on Windows, second realloc frees the block.
+  // Analyze 1: reported.
+  // Analyze 2: freed, irrelevant.
+  char* e3 = (char*)realloc(nullptr, 1023);
+  // e3 = (char*) realloc(e3, 0);
+  MOZ_ASSERT(e3);
+  Report(e3);
+
+  // Analyze 1: freed, irrelevant.
+  // Analyze 2: freed, irrelevant.
+  char* f1 = (char*)malloc(64);
+  UseItOrLoseIt(f1, aSeven);
+  free(f1);
+
+  // Analyze 1: ignored.
+  // Analyze 2: irrelevant.
+  Report((void*)(intptr_t)0x0);
+
+  // Analyze 1: mixture of reported and unreported.
+  // Analyze 2: all unreported.
+  Foo(aSeven);
+
+  // Analyze 1: twice-reported.
+  // Analyze 2: twice-reported.
+  char* g1 = (char*)malloc(77);
+  ReportOnAlloc(g1);
+  ReportOnAlloc(g1);
+
+  // Analyze 1: mixture of reported and unreported.
+  // Analyze 2: all unreported.
+  // Nb: this Foo() call is deliberately not adjacent to the previous one. See
+  // the comment about adjacent calls in Foo() for more details.
+  Foo(aSeven);
+
+  // Analyze 1: twice-reported.
+  // Analyze 2: once-reported.
+  char* g2 = (char*)malloc(78);
+  Report(g2);
+  ReportOnAlloc(g2);
+
+  // Analyze 1: twice-reported.
+  // Analyze 2: once-reported.
+  char* g3 = (char*)malloc(79);
+  ReportOnAlloc(g3);
+  Report(g3);
+
+  // All the odd-ball ones.
+  // Analyze 1: all unreported.
+  // Analyze 2: all freed, irrelevant.
+  // XXX: no memalign on Mac
+  // void* w = memalign(64, 65);           // rounds up to 128
+  // UseItOrLoseIt(w, aSeven);
+
+  // XXX: posix_memalign doesn't work on B2G
+  // void* x;
+  // posix_memalign(&y, 128, 129);         // rounds up to 256
+  // UseItOrLoseIt(x, aSeven);
+
+  // XXX: valloc doesn't work on Windows.
+  // void* y = valloc(1);                  // rounds up to 4096
+  // UseItOrLoseIt(y, aSeven);
+
+  // XXX: C11 only
+  // void* z = aligned_alloc(64, 256);
+  // UseItOrLoseIt(z, aSeven);
+
+  if (aNum == 1) {
+    // Analyze 1.
+    Analyze(std::move(f));
+  }
+
+  ClearReports();
+
+  //---------
+
+  Report(a2);
+  Report(a2);
+  free(c);
+  free(e);
+  Report(e2);
+  free(e3);
+  // free(w);
+  // free(x);
+  // free(y);
+  // free(z);
+
+  // Do some allocations that will only show up in cumulative mode.
+  for (int i = 0; i < 100; i++) {
+    void* v = malloc(128);
+    UseItOrLoseIt(v, aSeven);
+    free(v);
+  }
+
+  if (aNum == 2) {
+    // Analyze 2.
+    Analyze(std::move(f));
+  }
+}
+
+void TestPartial(const char* aTestName, const char* aMode, int aSeven) {
+  char filename[128];
+  sprintf(filename, "complete-%s-%s.json", aTestName, aMode);
+  auto f = MakeUnique<FpWriteFunc>(filename);
+
+  char options[128];
+  sprintf(options, "--mode=%s", aMode);
+  ResetEverything(options);
+
+  int kTenThousand = aSeven + 9993;
+  char* s;
+
+  // The output of this function is deterministic but it relies on the
+  // probability and seeds given to the FastBernoulliTrial instance in
+  // ResetBernoulli(). If they change, the output will change too.
+
+  // Expected fraction with stacks: (1 - (1 - 0.003) ** 16) = 0.0469.
+  // So we expect about 0.0469 * 10000 == 469.
+  // We actually get 511.
+  for (int i = 0; i < kTenThousand; i++) {
+    s = (char*)malloc(16);
+    UseItOrLoseIt(s, aSeven);
+  }
+
+  // Expected fraction with stacks: (1 - (1 - 0.003) ** 128) = 0.3193.
+  // So we expect about 0.3193 * 10000 == 3193.
+  // We actually get 3136.
+  for (int i = 0; i < kTenThousand; i++) {
+    s = (char*)malloc(128);
+    UseItOrLoseIt(s, aSeven);
+  }
+
+  // Expected fraction with stacks: (1 - (1 - 0.003) ** 1024) = 0.9539.
+  // So we expect about 0.9539 * 10000 == 9539.
+  // We actually get 9531.
+  for (int i = 0; i < kTenThousand; i++) {
+    s = (char*)malloc(1024);
+    UseItOrLoseIt(s, aSeven);
+  }
+
+  Analyze(std::move(f));
+}
+
+void TestScan(int aSeven) {
+  auto f = MakeUnique<FpWriteFunc>("basic-scan.json");
+
+  ResetEverything("--mode=scan");
+
+  uintptr_t* p = (uintptr_t*)malloc(6 * sizeof(uintptr_t));
+  UseItOrLoseIt(p, aSeven);
+
+  // Hard-coded values checked by scan-test.py
+  p[0] = 0x123;                         // outside a block, small value
+  p[1] = 0x0;                           // null
+  p[2] = (uintptr_t)((uint8_t*)p - 1);  // pointer outside a block, but nearby
+  p[3] = (uintptr_t)p;                  // pointer to start of a block
+  p[4] = (uintptr_t)((uint8_t*)p + 1);  // pointer into a block
+  p[5] = 0x0;                           // trailing null
+
+  Analyze(std::move(f));
+}
+
+void RunTests() {
+  // This test relies on the compiler not doing various optimizations, such as
+  // eliding unused malloc() calls or unrolling loops with fixed iteration
+  // counts. So we compile it with -O0 (or equivalent), which probably prevents
+  // that. We also use the following variable for various loop iteration
+  // counts, just in case compilers might unroll very small loops even with
+  // -O0.
+  int seven = 7;
+
+  // Make sure that DMD is actually running; it is initialized on the first
+  // allocation.
+  int* x = (int*)malloc(100);
+  UseItOrLoseIt(x, seven);
+  MOZ_RELEASE_ASSERT(IsRunning());
+
+  // Please keep this in sync with run_test in test_dmd.js.
+
+  TestEmpty("empty", "live");
+  TestEmpty("empty", "dark-matter");
+  TestEmpty("empty", "cumulative");
+
+  TestFull("full", 1, "live", seven);
+  TestFull("full", 1, "dark-matter", seven);
+
+  TestFull("full", 2, "dark-matter", seven);
+  TestFull("full", 2, "cumulative", seven);
+
+  TestPartial("partial", "live", seven);
+
+  TestScan(seven);
+}
+
+int main() {
+  RunTests();
+
+  return 0;
+}
diff --git a/memory/replace/dmd/test/basic-scan-32-expected.txt b/memory/replace/dmd/test/basic-scan-32-expected.txt
new file mode 100644
index 0000000000..9f6f4db325
--- /dev/null
+++ b/memory/replace/dmd/test/basic-scan-32-expected.txt
@@ -0,0 +1,25 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o basic-scan-32-actual.txt --clamp-contents basic-scan.json
+
+Invocation {
+  $DMD = '--mode=scan'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 1
+  32 bytes (24 requested / 8 slop)
+  100.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 32 bytes in 1 blocks
+}
+
diff --git a/memory/replace/dmd/test/basic-scan-64-expected.txt b/memory/replace/dmd/test/basic-scan-64-expected.txt
new file mode 100644
index 0000000000..59effc07b7
--- /dev/null
+++ b/memory/replace/dmd/test/basic-scan-64-expected.txt
@@ -0,0 +1,25 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o basic-scan-64-actual.txt --clamp-contents basic-scan.json
+
+Invocation {
+  $DMD = '--mode=scan'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 1
+  48 bytes (48 requested / 0 slop)
+  100.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 48 bytes in 1 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-empty-cumulative-expected.txt b/memory/replace/dmd/test/complete-empty-cumulative-expected.txt
new file mode 100644
index 0000000000..2486015d0b
--- /dev/null
+++ b/memory/replace/dmd/test/complete-empty-cumulative-expected.txt
@@ -0,0 +1,18 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-empty-cumulative-actual.txt complete-empty-cumulative.json
+
+Invocation {
+  $DMD = '--mode=cumulative --stacks=full'
+  Mode = 'cumulative'
+}
+
+#-----------------------------------------------------------------
+
+# no cumulative heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 0 bytes in 0 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-empty-dark-matter-expected.txt b/memory/replace/dmd/test/complete-empty-dark-matter-expected.txt
new file mode 100644
index 0000000000..0020cddde3
--- /dev/null
+++ b/memory/replace/dmd/test/complete-empty-dark-matter-expected.txt
@@ -0,0 +1,29 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-empty-dark-matter-actual.txt complete-empty-dark-matter.json
+
+Invocation {
+  $DMD = '--mode=dark-matter --stacks=full'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+# no twice-reported heap blocks
+
+#-----------------------------------------------------------------
+
+# no unreported heap blocks
+
+#-----------------------------------------------------------------
+
+# no once-reported heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                     0 bytes (100.00%) in       0 blocks (100.00%)
+  Unreported:                0 bytes (  0.00%) in       0 blocks (  0.00%)
+  Once-reported:             0 bytes (  0.00%) in       0 blocks (  0.00%)
+  Twice-reported:            0 bytes (  0.00%) in       0 blocks (  0.00%)
+}
+
diff --git a/memory/replace/dmd/test/complete-empty-live-expected.txt b/memory/replace/dmd/test/complete-empty-live-expected.txt
new file mode 100644
index 0000000000..d0d1721965
--- /dev/null
+++ b/memory/replace/dmd/test/complete-empty-live-expected.txt
@@ -0,0 +1,18 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-empty-live-actual.txt complete-empty-live.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+# no live heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 0 bytes in 0 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-full1-dark-matter-expected.txt b/memory/replace/dmd/test/complete-full1-dark-matter-expected.txt
new file mode 100644
index 0000000000..2c7d6b6343
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full1-dark-matter-expected.txt
@@ -0,0 +1,265 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full1-dark-matter-actual.txt complete-full1-dark-matter.json
+
+Invocation {
+  $DMD = '--mode=dark-matter --stacks=full --show-dump-stats=yes'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+Twice-reported {
+  1 block in heap block record 1 of 4
+  80 bytes (79 requested / 1 slop)
+  0.66% of the heap (0.66% cumulative)
+  29.41% of twice-reported (29.41% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 2 of 4
+  80 bytes (78 requested / 2 slop)
+  0.66% of the heap (1.32% cumulative)
+  29.41% of twice-reported (58.82% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 3 of 4
+  80 bytes (77 requested / 3 slop)
+  0.66% of the heap (1.98% cumulative)
+  29.41% of twice-reported (88.24% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 4 of 4
+  32 bytes (30 requested / 2 slop)
+  0.26% of the heap (2.25% cumulative)
+  11.76% of twice-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Unreported {
+  9 blocks in heap block record 1 of 3
+  1,008 bytes (900 requested / 108 slop)
+  Individual block sizes: 112 x 9
+  8.33% of the heap (8.33% cumulative)
+  81.82% of unreported (81.82% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  2 blocks in heap block record 2 of 3
+  112 bytes (112 requested / 0 slop)
+  Individual block sizes: 64; 48
+  0.93% of the heap (9.26% cumulative)
+  9.09% of unreported (90.91% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  2 blocks in heap block record 3 of 3
+  112 bytes (112 requested / 0 slop)
+  Individual block sizes: 64; 48
+  0.93% of the heap (10.19% cumulative)
+  9.09% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Once-reported {
+  1 block in heap block record 1 of 11
+  8,192 bytes (7,169 requested / 1,023 slop)
+  67.72% of the heap (67.72% cumulative)
+  77.34% of once-reported (77.34% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 2 of 11
+  1,024 bytes (1,023 requested / 1 slop)
+  8.47% of the heap (76.19% cumulative)
+  9.67% of once-reported (87.01% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 3 of 11
+  512 bytes (512 requested / 0 slop)
+  4.23% of the heap (80.42% cumulative)
+  4.83% of once-reported (91.84% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  2 blocks in heap block record 4 of 11
+  240 bytes (240 requested / 0 slop)
+  Individual block sizes: 128; 112
+  1.98% of the heap (82.41% cumulative)
+  2.27% of once-reported (94.11% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  2 blocks in heap block record 5 of 11
+  240 bytes (240 requested / 0 slop)
+  Individual block sizes: 128; 112
+  1.98% of the heap (84.39% cumulative)
+  2.27% of once-reported (96.37% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 6 of 11
+  96 bytes (96 requested / 0 slop)
+  0.79% of the heap (85.19% cumulative)
+  0.91% of once-reported (97.28% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 7 of 11
+  96 bytes (96 requested / 0 slop)
+  0.79% of the heap (85.98% cumulative)
+  0.91% of once-reported (98.19% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 8 of 11
+  80 bytes (80 requested / 0 slop)
+  0.66% of the heap (86.64% cumulative)
+  0.76% of once-reported (98.94% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 9 of 11
+  80 bytes (80 requested / 0 slop)
+  0.66% of the heap (87.30% cumulative)
+  0.76% of once-reported (99.70% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 10 of 11
+  16 bytes (16 requested / 0 slop)
+  0.13% of the heap (87.43% cumulative)
+  0.15% of once-reported (99.85% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 11 of 11
+  16 bytes (10 requested / 6 slop)
+  0.13% of the heap (87.57% cumulative)
+  0.15% of once-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                12,096 bytes (100.00%) in      30 blocks (100.00%)
+  Unreported:            1,232 bytes ( 10.19%) in      13 blocks ( 43.33%)
+  Once-reported:        10,592 bytes ( 87.57%) in      13 blocks ( 43.33%)
+  Twice-reported:          272 bytes (  2.25%) in       4 blocks ( 13.33%)
+}
+
diff --git a/memory/replace/dmd/test/complete-full1-live-expected.txt b/memory/replace/dmd/test/complete-full1-live-expected.txt
new file mode 100644
index 0000000000..eaa1883e1f
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full1-live-expected.txt
@@ -0,0 +1,127 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full1-live-actual.txt complete-full1-live.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full --show-dump-stats=yes'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 12
+  8,192 bytes (7,169 requested / 1,023 slop)
+  67.72% of the heap (67.72% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 2 of 12
+  1,024 bytes (1,023 requested / 1 slop)
+  8.47% of the heap (76.19% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  9 blocks in heap block record 3 of 12
+  1,008 bytes (900 requested / 108 slop)
+  Individual block sizes: 112 x 9
+  8.33% of the heap (84.52% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  6 blocks in heap block record 4 of 12
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  4.37% of the heap (88.89% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  6 blocks in heap block record 5 of 12
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  4.37% of the heap (93.25% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 6 of 12
+  512 bytes (512 requested / 0 slop)
+  4.23% of the heap (97.49% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 7 of 12
+  80 bytes (79 requested / 1 slop)
+  0.66% of the heap (98.15% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 8 of 12
+  80 bytes (78 requested / 2 slop)
+  0.66% of the heap (98.81% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 9 of 12
+  80 bytes (77 requested / 3 slop)
+  0.66% of the heap (99.47% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 10 of 12
+  32 bytes (30 requested / 2 slop)
+  0.26% of the heap (99.74% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 11 of 12
+  16 bytes (16 requested / 0 slop)
+  0.13% of the heap (99.87% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 12 of 12
+  16 bytes (10 requested / 6 slop)
+  0.13% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 12,096 bytes in 30 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-full2-cumulative-expected.txt b/memory/replace/dmd/test/complete-full2-cumulative-expected.txt
new file mode 100644
index 0000000000..5a225b9b8e
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full2-cumulative-expected.txt
@@ -0,0 +1,173 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full2-cumulative-actual.txt complete-full2-cumulative.json
+
+Invocation {
+  $DMD = '--mode=cumulative --stacks=full --show-dump-stats=yes'
+  Mode = 'cumulative'
+}
+
+#-----------------------------------------------------------------
+
+Cumulative {
+  100 blocks in heap block record 1 of 17
+  12,800 bytes (12,800 requested / 0 slop)
+  Individual block sizes: 128 x 100
+  42.37% of the heap (42.37% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 2 of 17
+  8,192 bytes (7,169 requested / 1,023 slop)
+  27.12% of the heap (69.49% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 3 of 17
+  4,096 bytes (4,096 requested / 0 slop)
+  13.56% of the heap (83.05% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  10 blocks in heap block record 4 of 17
+  1,120 bytes (1,000 requested / 120 slop)
+  Individual block sizes: 112 x 10
+  3.71% of the heap (86.76% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 5 of 17
+  1,024 bytes (1,024 requested / 0 slop)
+  3.39% of the heap (90.15% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 6 of 17
+  1,024 bytes (1,023 requested / 1 slop)
+  3.39% of the heap (93.54% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  6 blocks in heap block record 7 of 17
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  1.75% of the heap (95.29% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  6 blocks in heap block record 8 of 17
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  1.75% of the heap (97.03% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 9 of 17
+  512 bytes (512 requested / 0 slop)
+  1.69% of the heap (98.73% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 10 of 17
+  80 bytes (79 requested / 1 slop)
+  0.26% of the heap (98.99% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 11 of 17
+  80 bytes (78 requested / 2 slop)
+  0.26% of the heap (99.26% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 12 of 17
+  80 bytes (77 requested / 3 slop)
+  0.26% of the heap (99.52% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 13 of 17
+  64 bytes (64 requested / 0 slop)
+  0.21% of the heap (99.74% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 14 of 17
+  32 bytes (30 requested / 2 slop)
+  0.11% of the heap (99.84% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 15 of 17
+  16 bytes (16 requested / 0 slop)
+  0.05% of the heap (99.89% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 16 of 17
+  16 bytes (16 requested / 0 slop)
+  0.05% of the heap (99.95% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 17 of 17
+  16 bytes (10 requested / 6 slop)
+  0.05% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 30,208 bytes in 135 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-full2-dark-matter-expected.txt b/memory/replace/dmd/test/complete-full2-dark-matter-expected.txt
new file mode 100644
index 0000000000..5f9585a8c6
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full2-dark-matter-expected.txt
@@ -0,0 +1,140 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full2-dark-matter-actual.txt complete-full2-dark-matter.json
+
+Invocation {
+  $DMD = '--mode=dark-matter --stacks=full --show-dump-stats=yes'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+Twice-reported {
+  1 block in heap block record 1 of 2
+  80 bytes (77 requested / 3 slop)
+  2.81% of the heap (2.81% cumulative)
+  83.33% of twice-reported (83.33% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 2 of 2
+  16 bytes (16 requested / 0 slop)
+  0.56% of the heap (3.37% cumulative)
+  16.67% of twice-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Unreported {
+  9 blocks in heap block record 1 of 3
+  1,008 bytes (900 requested / 108 slop)
+  Individual block sizes: 112 x 9
+  35.39% of the heap (35.39% cumulative)
+  48.84% of unreported (48.84% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  6 blocks in heap block record 2 of 3
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  18.54% of the heap (53.93% cumulative)
+  25.58% of unreported (74.42% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  6 blocks in heap block record 3 of 3
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  18.54% of the heap (72.47% cumulative)
+  25.58% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Once-reported {
+  1 block in heap block record 1 of 4
+  512 bytes (512 requested / 0 slop)
+  17.98% of the heap (17.98% cumulative)
+  74.42% of once-reported (74.42% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 2 of 4
+  80 bytes (79 requested / 1 slop)
+  2.81% of the heap (20.79% cumulative)
+  11.63% of once-reported (86.05% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 3 of 4
+  80 bytes (78 requested / 2 slop)
+  2.81% of the heap (23.60% cumulative)
+  11.63% of once-reported (97.67% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 4 of 4
+  16 bytes (10 requested / 6 slop)
+  0.56% of the heap (24.16% cumulative)
+  2.33% of once-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                 2,848 bytes (100.00%) in      27 blocks (100.00%)
+  Unreported:            2,064 bytes ( 72.47%) in      21 blocks ( 77.78%)
+  Once-reported:           688 bytes ( 24.16%) in       4 blocks ( 14.81%)
+  Twice-reported:           96 bytes (  3.37%) in       2 blocks (  7.41%)
+}
+
diff --git a/memory/replace/dmd/test/complete-partial-live-expected.txt b/memory/replace/dmd/test/complete-partial-live-expected.txt
new file mode 100644
index 0000000000..e7f27b0ee6
--- /dev/null
+++ b/memory/replace/dmd/test/complete-partial-live-expected.txt
@@ -0,0 +1,56 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-partial-live-actual.txt complete-partial-live.json
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  9,531 blocks in heap block record 1 of 4
+  9,759,744 bytes (9,759,744 requested / 0 slop)
+  Individual block sizes: 1,024 x 9,531
+  83.56% of the heap (83.56% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  16,822 blocks in heap block record 2 of 4
+  1,510,672 bytes (1,510,672 requested / 0 slop)
+  Individual block sizes: 1,024 x 469; 128 x 6,864; 16 x 9,489
+  12.93% of the heap (96.49% cumulative)
+  Allocated at {
+    #01: (no stack trace recorded due to --stacks=partial)
+  }
+}
+
+Live {
+  3,136 blocks in heap block record 3 of 4
+  401,408 bytes (401,408 requested / 0 slop)
+  Individual block sizes: 128 x 3,136
+  3.44% of the heap (99.93% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  511 blocks in heap block record 4 of 4
+  8,176 bytes (8,176 requested / 0 slop)
+  Individual block sizes: 16 x 511
+  0.07% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 11,680,000 bytes in 30,000 blocks
+}
+
diff --git a/memory/replace/dmd/test/moz.build b/memory/replace/dmd/test/moz.build
new file mode 100644
index 0000000000..b869a86cc7
--- /dev/null
+++ b/memory/replace/dmd/test/moz.build
@@ -0,0 +1,29 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+GeckoSimplePrograms(
+    [
+        "SmokeDMD",
+    ],
+    linkage=None,
+)
+
+# See the comment at the top of SmokeDMD.cpp:RunTests().
+if CONFIG["CC_TYPE"] == "clang-cl":
+    CXXFLAGS += ["-Od", "-clang:-fno-lto"]
+else:
+    CXXFLAGS += ["-O0", "-fno-lto"]
+
+DEFINES["MOZ_NO_MOZALLOC"] = True
+
+DisableStlWrapping()
+
+XPCSHELL_TESTS_MANIFESTS += [
+    "xpcshell.ini",
+]
+
+if CONFIG["CC_TYPE"] in ("clang", "gcc"):
+    CXXFLAGS += ["-Wno-error=shadow"]
diff --git a/memory/replace/dmd/test/scan-test.py b/memory/replace/dmd/test/scan-test.py
new file mode 100644
index 0000000000..11cbfaa32c
--- /dev/null
+++ b/memory/replace/dmd/test/scan-test.py
@@ -0,0 +1,104 @@
+#! /usr/bin/env python
+#
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+"""Testing for the JSON file emitted by DMD heap scan mode when running SmokeDMD."""
+
+from __future__ import absolute_import, print_function, division
+
+import argparse
+import gzip
+import json
+import sys
+
+# The DMD output version this script handles.
+outputVersion = 5
+
+
+def parseCommandLine():
+    description = """
+Ensure that DMD heap scan mode creates the correct output when run with SmokeDMD.
+This is only for testing. Input files can be gzipped.
+"""
+    p = argparse.ArgumentParser(description=description)
+
+    p.add_argument(
+        "--clamp-contents",
+        action="store_true",
+        help="expect that the contents of the JSON input file have had "
+        "their addresses clamped",
+    )
+
+    p.add_argument("input_file", help="a file produced by DMD")
+
+    return p.parse_args(sys.argv[1:])
+
+
+def checkScanContents(contents, expected):
+    if len(contents) != len(expected):
+        raise Exception(
+            "Expected "
+            + str(len(expected))
+            + " things in contents but found "
+            + str(len(contents))
+        )
+
+    for i in range(len(expected)):
+        if contents[i] != expected[i]:
+            raise Exception(
+                "Expected to find "
+                + expected[i]
+                + " at offset "
+                + str(i)
+                + " but found "
+                + contents[i]
+            )
+
+
+def main():
+    args = parseCommandLine()
+
+    # Handle gzipped input if necessary.
+    isZipped = args.input_file.endswith(".gz")
+    opener = gzip.open if isZipped else open
+
+    with opener(args.input_file, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    invocation = j["invocation"]
+
+    mode = invocation["mode"]
+    if mode != "scan":
+        raise Exception("bad 'mode' property: '{:s}'".format(mode))
+
+    blockList = j["blockList"]
+
+    if len(blockList) != 1:
+        raise Exception("Expected only one block")
+
+    b = blockList[0]
+
+    # The expected values are based on hard-coded values in SmokeDMD.cpp.
+    if args.clamp_contents:
+        expected = ["0", "0", "0", b["addr"], b["addr"]]
+    else:
+        addr = int(b["addr"], 16)
+        expected = [
+            "123",
+            "0",
+            str(format(addr - 1, "x")),
+            b["addr"],
+            str(format(addr + 1, "x")),
+            "0",
+        ]
+
+    checkScanContents(b["contents"], expected)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/memory/replace/dmd/test/script-diff-dark-matter-expected.txt b/memory/replace/dmd/test/script-diff-dark-matter-expected.txt
new file mode 100644
index 0000000000..b1fc28bac5
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-dark-matter-expected.txt
@@ -0,0 +1,127 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-diff-dark-matter-actual.txt script-diff-dark-matter1.json script-diff-dark-matter2.json
+
+Invocation 1 {
+  $DMD = '--mode=dark-matter'
+  Mode = 'dark-matter'
+}
+
+Invocation 2 {
+  $DMD = '1'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+Twice-reported {
+  -1 blocks in heap block record 1 of 1
+  -1,088 bytes (-1,064 requested / -24 slop)
+  Individual block sizes: -1,024; -127; 63
+  15.46% of the heap (15.46% cumulative)
+  100.00% of twice-reported (100.00% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+  Reported at {
+    #01: R1 (R1.cpp:99)
+  }
+  Reported again at {
+    #01: R2 (R2.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Unreported {
+  4 blocks in heap block record 1 of 5
+  16,384 bytes (16,384 requested / 0 slop)
+  Individual block sizes: 4,096 x 4
+  -232.76% of the heap (-232.76% cumulative)
+  371.01% of unreported (371.01% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+Unreported {
+  7 blocks in heap block record 2 of 5
+  -11,968 bytes (-12,016 requested / 48 slop)
+  Individual block sizes: -15,360; 2,048; 512 x 2; 128; -127; 64 x 4; 63
+  170.02% of the heap (-62.74% cumulative)
+  -271.01% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+}
+
+Unreported {
+  0 blocks in heap block record 3 of 5
+  0 bytes (-384 requested / 384 slop)
+  Individual block sizes: (no change)
+  -0.00% of the heap (-62.74% cumulative)
+  0.00% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Unreported {
+  -2 blocks in heap block record 4 of 5
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 8,192 x 2; -4,096 x 4
+  -0.00% of the heap (-62.74% cumulative)
+  0.00% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Unreported {
+  0 blocks in heap block record 5 of 5
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 20,480; -16,384; -8,192; 4,096
+  -0.00% of the heap (-62.74% cumulative)
+  0.00% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: (no stack trace recorded due to --stacks=partial)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Once-reported {
+  -3 blocks in heap block record 1 of 2
+  -10,240 bytes (-10,192 requested / -48 slop)
+  Individual block sizes: -4,096 x 2; -2,048
+  145.48% of the heap (145.48% cumulative)
+  98.77% of once-reported (98.77% cumulative)
+  Allocated at {
+    #01: D (D.cpp:99)
+  }
+  Reported at {
+    #01: R1 (R1.cpp:99)
+  }
+}
+
+Once-reported {
+  -1 blocks in heap block record 2 of 2
+  -127 bytes (-151 requested / 24 slop)
+  1.80% of the heap (147.28% cumulative)
+  1.23% of once-reported (100.00% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+  Reported at {
+    #01: R1 (R1.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                -7,039 bytes (100.00%) in       4 blocks (100.00%)
+  Unreported:            4,416 bytes (-62.74%) in       9 blocks (225.00%)
+  Once-reported:       -10,367 bytes (147.28%) in      -4 blocks (-100.00%)
+  Twice-reported:       -1,088 bytes ( 15.46%) in      -1 blocks (-25.00%)
+}
+
diff --git a/memory/replace/dmd/test/script-diff-dark-matter1.json b/memory/replace/dmd/test/script-diff-dark-matter1.json
new file mode 100644
index 0000000000..c8edafebe9
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-dark-matter1.json
@@ -0,0 +1,51 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=dark-matter",
+  "mode": "dark-matter"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 4},
+
+  {"req": 4096, "alloc": "B", "num": 3},
+  {"req": 4096, "alloc": "B"},
+
+  {"req": 4096, "alloc": "C", "num": 2},
+  {"req": 4096, "alloc": "C", "num": 2},
+
+  {"req": 4096,             "alloc": "D", "reps": ["R1"], "num": 2},
+  {"req": 2000, "slop": 48, "alloc": "D", "reps": ["R1"]},
+
+  {"req": 15360,            "alloc": "F"},
+  {"req": 512,              "alloc": "F", "num": 2},
+  {"req": 127,              "alloc": "F"},
+  {"req": 1024,             "alloc": "F", "reps": ["R1"]},
+  {"req": 127,              "alloc": "F", "reps": ["R1"]},
+  {"req": 1000, "slop": 24, "alloc": "F", "reps": ["R1", "R2"]},
+  {"req": 127,              "alloc": "F", "reps": ["R1", "R2"]},
+
+  {"req": 4096 },
+  {"req": 8192 },
+  {"req": 16384 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-diff-dark-matter2.json b/memory/replace/dmd/test/script-diff-dark-matter2.json
new file mode 100644
index 0000000000..94b8888b89
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-dark-matter2.json
@@ -0,0 +1,51 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "1",
+  "mode": "dark-matter"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 4},
+
+  {"req": 8192, "alloc": "B"},
+  {"req": 8192, "alloc": "B"},
+
+  {"req": 4000, "slop": 96, "alloc": "C", "num": 4},
+
+  {"req": 4096, "alloc": "E", "num": 4},
+
+  {"req": 2000, "slop": 48, "alloc": "F"},
+  {"req": 1000, "slop": 24, "alloc": "F", "reps": ["R1"]},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 128,              "alloc": "F"},
+  {"req": 63,               "alloc": "F", "reps": ["R1", "R2"]},
+  {"req": 64,               "alloc": "F", "num": 4},
+  {"req": 63,               "alloc": "F"},
+
+  {"req": 4096, "num": 2 },
+  {"req": 20480 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-diff-live-expected.txt b/memory/replace/dmd/test/script-diff-live-expected.txt
new file mode 100644
index 0000000000..ecd291ad8c
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-live-expected.txt
@@ -0,0 +1,81 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-diff-live-actual.txt script-diff-live1.json script-diff-live2.json
+
+Invocation 1 {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+Invocation 2 {
+  $DMD = '--mode=live --stacks=partial'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  4 blocks in heap block record 1 of 6
+  16,384 bytes (16,384 requested / 0 slop)
+  Individual block sizes: 4,096 x 4
+  -232.76% of the heap (-232.76% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 6
+  -13,183 bytes (-13,231 requested / 48 slop)
+  Individual block sizes: -15,360; 2,048; -1,024; 512 x 2; 128; -127 x 3; 64 x 4; 63 x 2
+  187.29% of the heap (-45.48% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+}
+
+Live {
+  -3 blocks in heap block record 3 of 6
+  -10,240 bytes (-10,192 requested / -48 slop)
+  Individual block sizes: -4,096 x 2; -2,048
+  145.48% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: D (D.cpp:99)
+  }
+}
+
+Live {
+  0 blocks in heap block record 4 of 6
+  0 bytes (-384 requested / 384 slop)
+  Individual block sizes: (no change)
+  -0.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Live {
+  0 blocks in heap block record 5 of 6
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 20,480; -16,384; -8,192; 4,096
+  -0.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: (no stack trace recorded due to --stacks=partial)
+  }
+}
+
+Live {
+  -2 blocks in heap block record 6 of 6
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 8,192 x 2; -4,096 x 4
+  -0.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: -7,039 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-diff-live1.json b/memory/replace/dmd/test/script-diff-live1.json
new file mode 100644
index 0000000000..87e07aed5e
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-live1.json
@@ -0,0 +1,51 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=live",
+  "mode": "live"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 4},
+
+  {"req": 4096, "alloc": "B", "num": 4},
+
+  {"req": 4096, "alloc": "C", "num": 4},
+
+  {"req": 4096,             "alloc": "D"},
+  {"req": 4096,             "alloc": "D"},
+  {"req": 2000, "slop": 48, "alloc": "D"},
+
+  {"req": 15360,            "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 127,              "alloc": "F"},
+  {"req": 1024,             "alloc": "F"},
+  {"req": 127,              "alloc": "F"},
+  {"req": 1000, "slop": 24, "alloc": "F"},
+  {"req": 127,              "alloc": "F"},
+
+  {"req": 4096 },
+  {"req": 8192 },
+  {"req": 16384 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-diff-live2.json b/memory/replace/dmd/test/script-diff-live2.json
new file mode 100644
index 0000000000..4c7476f4c3
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-live2.json
@@ -0,0 +1,53 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=live --stacks=partial",
+  "mode": "live"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 3},
+  {"req": 4096, "alloc": "A"},
+
+  {"req": 8192, "alloc": "B"},
+  {"req": 8192, "alloc": "B"},
+
+  {"req": 4000, "slop": 96, "alloc": "C", "num": 4},
+
+  {"req": 4096, "alloc": "E"},
+  {"req": 4096, "alloc": "E"},
+  {"req": 4096, "alloc": "E"},
+  {"req": 4096, "alloc": "E"},
+
+  {"req": 2000, "slop": 48, "alloc": "F"},
+  {"req": 1000, "slop": 24, "alloc": "F"},
+  {"req": 512,              "alloc": "F", "num": 4},
+  {"req": 128,              "alloc": "F"},
+  {"req": 63,               "alloc": "F"},
+  {"req": 64,               "alloc": "F", "num": 4},
+  {"req": 63,               "alloc": "F"},
+
+  {"req": 4096 },
+  {"req": 4096 },
+  {"req": 20480 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-ignore-alloc-fns-expected.txt b/memory/replace/dmd/test/script-ignore-alloc-fns-expected.txt
new file mode 100644
index 0000000000..9428ef45fb
--- /dev/null
+++ b/memory/replace/dmd/test/script-ignore-alloc-fns-expected.txt
@@ -0,0 +1,72 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-ignore-alloc-fns-actual.txt --ignore-alloc-fns script-ignore-alloc-fns.json
+
+Invocation {
+  $DMD = '1'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+# no twice-reported heap blocks
+
+#-----------------------------------------------------------------
+
+Unreported {
+  1 block in heap block record 1 of 4
+  1,048,576 bytes (1,048,576 requested / 0 slop)
+  93.22% of the heap (93.22% cumulative)
+  93.22% of unreported (93.22% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+Unreported {
+  1 block in heap block record 2 of 4
+  65,536 bytes (65,536 requested / 0 slop)
+  5.83% of the heap (99.05% cumulative)
+  5.83% of unreported (99.05% cumulative)
+  Allocated at {
+    #01: js::jit::JitRuntime::initialize(JSContext*) (Ion.cpp:301)
+  }
+}
+
+Unreported {
+  1 block in heap block record 3 of 4
+  8,192 bytes (8,000 requested / 192 slop)
+  0.73% of the heap (99.78% cumulative)
+  0.73% of unreported (99.78% cumulative)
+  Allocated at {
+    #01: mozilla::Vector::growStorageBy(unsigned long) (Vector.h:802)
+    #02: D (D.cpp:99)
+  }
+}
+
+Unreported {
+  1 block in heap block record 4 of 4
+  2,500 bytes (2,500 requested / 0 slop)
+  0.22% of the heap (100.00% cumulative)
+  0.22% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: g_type_create_instance (/usr/lib/x86_64-linux-gnu/libgobject-2.0.so.0)
+    #02: not_an_alloc_function_so_alloc_functions_below_here_will_not_be_stripped (blah)
+    #03: replace_posix_memalign (replace_malloc.h:120)
+    #04: ??? (/lib/x86_64-linux-gnu/libglib-2.0.so.0)
+    #05: another_non_alloc_function (blah)
+  }
+}
+
+#-----------------------------------------------------------------
+
+# no once-reported heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:             1,124,804 bytes (100.00%) in       4 blocks (100.00%)
+  Unreported:        1,124,804 bytes (100.00%) in       4 blocks (100.00%)
+  Once-reported:             0 bytes (  0.00%) in       0 blocks (  0.00%)
+  Twice-reported:            0 bytes (  0.00%) in       0 blocks (  0.00%)
+}
+
diff --git a/memory/replace/dmd/test/script-ignore-alloc-fns.json b/memory/replace/dmd/test/script-ignore-alloc-fns.json
new file mode 100644
index 0000000000..900d33e759
--- /dev/null
+++ b/memory/replace/dmd/test/script-ignore-alloc-fns.json
@@ -0,0 +1,46 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "1",
+  "mode": "dark-matter"
+ },
+ "blockList": [
+  {"req": 1048576,           "alloc": "A"},
+  {"req": 65536,             "alloc": "B"},
+  {"req": 8000, "slop": 192, "alloc": "C"},
+  {"req": 2500,              "alloc": "D"}
+ ],
+ "traceTable": {
+  "A": ["AA", "AB", "AC", "AD"],
+  "B": ["BA", "BB", "BC"],
+  "C": ["CA", "CB", "CC", "CD"],
+  "D": ["DA", "DB", "DD", "DD", "DE", "DF", "DG", "DH", "DI", "DJ"]
+ },
+ "frameTable": {
+  "AA": "#00: replace_malloc (DMD.cpp:1106)",
+  "AB": "#00: moz_xmalloc (mozalloc.cpp:68)",
+  "AC": "#00: operator new(unsigned long) (mozalloc.h:208)",
+  "AD": "#00: A (A.cpp:99)",
+
+  "BA": "#00: replace_calloc (DMD.cpp:1125)",
+  "BB": "#00: js_calloc(unsigned long) (Utility.h:107)",
+  "BC": "#06: js::jit::JitRuntime::initialize(JSContext*) (Ion.cpp:301)",
+
+  "CA": "#00: replace_realloc (DMD.cpp:1153)",
+  "CB": "#00: bool* mozilla::MallocAllocPolicy::pod_realloc<bool>(bool*, unsigned long, unsigned long) (AllocPolicy.h:74)",
+  "CC": "#00: mozilla::Vector::growStorageBy(unsigned long) (Vector.h:802)",
+  "CD": "#00: D (D.cpp:99)",
+
+  "DA": "#00: replace_memalign (DMD.cpp:1181)",
+  "DB": "#00: replace_posix_memalign (replace_malloc.h:120)",
+  "DC": "#00: ??? (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DD": "#00: g_slice_alloc (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DE": "#00: g_slice_alloc0 (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DF": "#00: g_type_create_instance (/usr/lib/x86_64-linux-gnu/libgobject-2.0.so.0)",
+  "DG": "#00: not_an_alloc_function_so_alloc_functions_below_here_will_not_be_stripped (blah)",
+  "DH": "#00: replace_posix_memalign (replace_malloc.h:120)",
+  "DI": "#00: ??? (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DJ": "#00: another_non_alloc_function (blah)"
+ }
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames-1-expected.txt b/memory/replace/dmd/test/script-max-frames-1-expected.txt
new file mode 100644
index 0000000000..65a00762bb
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames-1-expected.txt
@@ -0,0 +1,26 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-max-frames-1-actual.txt --max-frames=1 script-max-frames.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  4 blocks in heap block record 1 of 1
+  4,416 bytes (4,404 requested / 12 slop)
+  Individual block sizes: 4,096; 128; 112; 80
+  100.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 4,416 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames-3-expected.txt b/memory/replace/dmd/test/script-max-frames-3-expected.txt
new file mode 100644
index 0000000000..5df4914738
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames-3-expected.txt
@@ -0,0 +1,48 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-max-frames-3-actual.txt --max-frames=3 --no-fix-stacks script-max-frames.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  2 blocks in heap block record 1 of 3
+  4,224 bytes (4,224 requested / 0 slop)
+  Individual block sizes: 4,096; 128
+  95.65% of the heap (95.65% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: F (F.cpp:99)
+    #03: G (G.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 2 of 3
+  112 bytes (100 requested / 12 slop)
+  2.54% of the heap (98.19% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: X (X.cpp:99)
+    #03: Y (Y.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 3 of 3
+  80 bytes (80 requested / 0 slop)
+  1.81% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 4,416 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames-8-expected.txt b/memory/replace/dmd/test/script-max-frames-8-expected.txt
new file mode 100644
index 0000000000..174992d5b8
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames-8-expected.txt
@@ -0,0 +1,69 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-max-frames-8-actual.txt script-max-frames.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 4
+  4,096 bytes (4,096 requested / 0 slop)
+  92.75% of the heap (92.75% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: F (F.cpp:99)
+    #03: G (G.cpp:99)
+    #04: H (H.cpp:99)
+    #05: I (I.cpp:99)
+    #06: J (J.cpp:99)
+    #07: K (K.cpp:99)
+    #08: L (L.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 2 of 4
+  128 bytes (128 requested / 0 slop)
+  2.90% of the heap (95.65% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: F (F.cpp:99)
+    #03: G (G.cpp:99)
+    #04: R (R.cpp:99)
+    #05: S (S.cpp:99)
+    #06: T (T.cpp:99)
+    #07: U (U.cpp:99)
+    #08: V (V.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 3 of 4
+  112 bytes (100 requested / 12 slop)
+  2.54% of the heap (98.19% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: X (X.cpp:99)
+    #03: Y (Y.cpp:99)
+    #04: Z (Z.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 4 of 4
+  80 bytes (80 requested / 0 slop)
+  1.81% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 4,416 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames.json b/memory/replace/dmd/test/script-max-frames.json
new file mode 100644
index 0000000000..690d50fa7e
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames.json
@@ -0,0 +1,43 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=live --stacks=full",
+  "mode": "live"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A"},
+  {"req": 128, "alloc": "B"},
+  {"req": 100, "slop":12, "alloc": "C"},
+  {"req": 80, "alloc": "D"}
+ ],
+ "traceTable": {
+  "A": ["E", "F", "G", "H", "I", "J", "K", "L", "M", "N", "O", "P"],
+  "B": ["E", "F", "G", "R", "S", "T", "U", "V"],
+  "C": ["E", "X", "Y", "Z"],
+  "D": ["E"]
+ },
+ "frameTable": {
+  "E": "#00: E (E.cpp:99)",
+  "F": "#00: F (F.cpp:99)",
+  "G": "#00: G (G.cpp:99)",
+  "H": "#00: H (H.cpp:99)",
+  "I": "#00: I (I.cpp:99)",
+  "J": "#00: J (J.cpp:99)",
+  "K": "#00: K (K.cpp:99)",
+  "L": "#00: L (L.cpp:99)",
+  "M": "#00: M (M.cpp:99)",
+  "N": "#00: N (N.cpp:99)",
+  "O": "#00: O (O.cpp:99)",
+  "P": "#00: P (P.cpp:99)",
+  "Q": "#00: Q (Q.cpp:99)",
+  "R": "#00: R (R.cpp:99)",
+  "S": "#00: S (S.cpp:99)",
+  "T": "#00: T (T.cpp:99)",
+  "U": "#00: U (U.cpp:99)",
+  "V": "#00: V (V.cpp:99)",
+  "W": "#00: W (W.cpp:99)",
+  "X": "#00: X (X.cpp:99)",
+  "Y": "#00: Y (Y.cpp:99)",
+  "Z": "#00: Z (Z.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-sort-by-num-blocks-expected.txt b/memory/replace/dmd/test/script-sort-by-num-blocks-expected.txt
new file mode 100644
index 0000000000..8de03d953b
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-num-blocks-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-num-blocks-actual.txt --sort-by=num-blocks script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  8 blocks in heap block record 1 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (33.32% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (66.67% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 3 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by-req-expected.txt b/memory/replace/dmd/test/script-sort-by-req-expected.txt
new file mode 100644
index 0000000000..3ab21ba8f7
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-req-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-req-actual.txt --sort-by=req --no-fix-stacks script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  5 blocks in heap block record 1 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (33.33% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (66.68% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  8 blocks in heap block record 3 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by-slop-expected.txt b/memory/replace/dmd/test/script-sort-by-slop-expected.txt
new file mode 100644
index 0000000000..c325c7ed40
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-slop-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-slop-actual.txt --sort-by=slop script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  8 blocks in heap block record 1 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (33.32% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (66.67% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 3 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by-usable-expected.txt b/memory/replace/dmd/test/script-sort-by-usable-expected.txt
new file mode 100644
index 0000000000..8239a4759e
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-usable-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-usable-actual.txt --sort-by=usable script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  5 blocks in heap block record 1 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (33.35% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (66.68% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+Live {
+  8 blocks in heap block record 3 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by.json.gz b/memory/replace/dmd/test/script-sort-by.json.gz
new file mode 100644
index 0000000000..b2308bab40
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by.json.gz
diff --git a/memory/replace/dmd/test/test_dmd.js b/memory/replace/dmd/test/test_dmd.js
new file mode 100644
index 0000000000..ec7ee49bfa
--- /dev/null
+++ b/memory/replace/dmd/test/test_dmd.js
@@ -0,0 +1,228 @@
+/* -*- indent-tabs-mode: nil; js-indent-level: 2 -*-*/
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this file,
+ * You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+"use strict";
+
+const { FileUtils } = ChromeUtils.import(
+  "resource://gre/modules/FileUtils.jsm"
+);
+const { Services } = ChromeUtils.import("resource://gre/modules/Services.jsm");
+
+// The xpcshell test harness sets PYTHON so we can read it here.
+var gEnv = Cc["@mozilla.org/process/environment;1"].getService(
+  Ci.nsIEnvironment
+);
+var gPythonName = gEnv.get("PYTHON");
+
+// If we're testing locally, the executable file is in "CurProcD". Otherwise,
+// it is in another location that we have to find.
+function getExecutable(aFilename) {
+  let file = FileUtils.getFile("CurProcD", [aFilename]);
+  if (!file.exists()) {
+    file = FileUtils.getFile("CurWorkD", []);
+    while (file.path.includes("xpcshell")) {
+      file = file.parent;
+    }
+    file.append("bin");
+    file.append(aFilename);
+  }
+  return file;
+}
+
+var gIsWindows = Services.appinfo.OS === "WINNT";
+var gDmdTestFile = getExecutable("SmokeDMD" + (gIsWindows ? ".exe" : ""));
+
+var gDmdScriptFile = getExecutable("dmd.py");
+
+var gScanTestFile = FileUtils.getFile("CurWorkD", ["scan-test.py"]);
+
+function readFile(aFile) {
+  let fstream = Cc["@mozilla.org/network/file-input-stream;1"].createInstance(
+    Ci.nsIFileInputStream
+  );
+  let cstream = Cc["@mozilla.org/intl/converter-input-stream;1"].createInstance(
+    Ci.nsIConverterInputStream
+  );
+  fstream.init(aFile, -1, 0, 0);
+  cstream.init(fstream, "UTF-8", 0, 0);
+
+  let data = "";
+  let str = {};
+  let read = 0;
+  do {
+    // Read as much as we can and put it in str.value.
+    read = cstream.readString(0xffffffff, str);
+    data += str.value;
+  } while (read != 0);
+
+  cstream.close(); // this closes fstream
+  return data.replace(/\r/g, ""); // normalize line endings
+}
+
+function runProcess(aExeFile, aArgs) {
+  let process = Cc["@mozilla.org/process/util;1"].createInstance(Ci.nsIProcess);
+  process.init(aExeFile);
+  process.run(/* blocking = */ true, aArgs, aArgs.length);
+  return process.exitValue;
+}
+
+function test(aPrefix, aArgs) {
+  // DMD writes the JSON files to CurWorkD, so we do likewise here with
+  // |actualFile| for consistency. It is removed once we've finished.
+  let expectedFile = FileUtils.getFile("CurWorkD", [aPrefix + "-expected.txt"]);
+  let actualFile = FileUtils.getFile("CurWorkD", [aPrefix + "-actual.txt"]);
+
+  // Run dmd.py on the JSON file, producing |actualFile|.
+
+  let args = [
+    gDmdScriptFile.path,
+    "--filter-stacks-for-testing",
+    "-o",
+    actualFile.path,
+  ].concat(aArgs);
+
+  runProcess(new FileUtils.File(gPythonName), args);
+
+  // Compare |expectedFile| with |actualFile|. We produce nice diffs with
+  // /usr/bin/diff on systems that have it (Mac and Linux). Otherwise (Windows)
+  // we do a string compare of the file contents and then print them both if
+  // they don't match.
+
+  let success;
+  try {
+    let rv = runProcess(new FileUtils.File("/usr/bin/diff"), [
+      "-u",
+      expectedFile.path,
+      actualFile.path,
+    ]);
+    success = rv == 0;
+  } catch (e) {
+    let expectedData = readFile(expectedFile);
+    let actualData = readFile(actualFile);
+    success = expectedData === actualData;
+    if (!success) {
+      expectedData = expectedData.split("\n");
+      actualData = actualData.split("\n");
+      for (let i = 0; i < expectedData.length; i++) {
+        print("EXPECTED:" + expectedData[i]);
+      }
+      for (let i = 0; i < actualData.length; i++) {
+        print("  ACTUAL:" + actualData[i]);
+      }
+    }
+  }
+
+  ok(success, aPrefix);
+
+  actualFile.remove(true);
+}
+
+// Run scan-test.py on the JSON file and see if it succeeds.
+function scanTest(aJsonFilePath, aExtraArgs) {
+  let args = [gScanTestFile.path, aJsonFilePath].concat(aExtraArgs);
+
+  return runProcess(new FileUtils.File(gPythonName), args) == 0;
+}
+
+function run_test() {
+  let jsonFile, jsonFile2;
+
+  // These tests do complete end-to-end testing of DMD, i.e. both the C++ code
+  // that generates the JSON output, and the script that post-processes that
+  // output.
+  //
+  // Run these synchronously, because test() updates the complete*.json files
+  // in-place (to fix stacks) when it runs dmd.py, and that's not safe to do
+  // asynchronously.
+
+  gEnv.set("DMD", "1");
+
+  runProcess(gDmdTestFile, []);
+
+  function test2(aTestName, aMode) {
+    let name = "complete-" + aTestName + "-" + aMode;
+    jsonFile = FileUtils.getFile("CurWorkD", [name + ".json"]);
+    test(name, [jsonFile.path]);
+    jsonFile.remove(true);
+  }
+
+  // Please keep this in sync with RunTests() in SmokeDMD.cpp.
+
+  test2("empty", "live");
+  test2("empty", "dark-matter");
+  test2("empty", "cumulative");
+
+  test2("full1", "live");
+  test2("full1", "dark-matter");
+
+  test2("full2", "dark-matter");
+  test2("full2", "cumulative");
+
+  test2("partial", "live");
+
+  // Heap scan testing.
+  jsonFile = FileUtils.getFile("CurWorkD", ["basic-scan.json"]);
+  ok(scanTest(jsonFile.path), "Basic scan test");
+
+  let is64Bit = Services.appinfo.is64Bit;
+  let basicScanFileName = "basic-scan-" + (is64Bit ? "64" : "32");
+  test(basicScanFileName, ["--clamp-contents", jsonFile.path]);
+  ok(
+    scanTest(jsonFile.path, ["--clamp-contents"]),
+    "Scan with address clamping"
+  );
+
+  // Run the generic test a second time to ensure that the first time produced
+  // valid JSON output. "--clamp-contents" is passed in so we don't have to have
+  // more variants of the files.
+  test(basicScanFileName, ["--clamp-contents", jsonFile.path]);
+  jsonFile.remove(true);
+
+  // These tests only test the post-processing script. They use hand-written
+  // JSON files as input. Ideally the JSON files would contain comments
+  // explaining how they work, but JSON doesn't allow comments, so I've put
+  // explanations here.
+
+  // This just tests that stack traces of various lengths are truncated
+  // appropriately. The number of records in the output is different for each
+  // of the tested values.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-max-frames.json"]);
+  test("script-max-frames-8", [jsonFile.path]); // --max-frames=8 is the default
+  test("script-max-frames-3", [
+    "--max-frames=3",
+    "--no-fix-stacks",
+    jsonFile.path,
+  ]);
+  test("script-max-frames-1", ["--max-frames=1", jsonFile.path]);
+
+  // This file has three records that are shown in a different order for each
+  // of the different sort values. It also tests the handling of gzipped JSON
+  // files.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-sort-by.json.gz"]);
+  test("script-sort-by-usable", ["--sort-by=usable", jsonFile.path]);
+  test("script-sort-by-req", [
+    "--sort-by=req",
+    "--no-fix-stacks",
+    jsonFile.path,
+  ]);
+  test("script-sort-by-slop", ["--sort-by=slop", jsonFile.path]);
+  test("script-sort-by-num-blocks", ["--sort-by=num-blocks", jsonFile.path]);
+
+  // This file has several real stack traces taken from Firefox execution, each
+  // of which tests a different allocator function (or functions).
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-ignore-alloc-fns.json"]);
+  test("script-ignore-alloc-fns", ["--ignore-alloc-fns", jsonFile.path]);
+
+  // This tests "live"-mode diffs.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-diff-live1.json"]);
+  jsonFile2 = FileUtils.getFile("CurWorkD", ["script-diff-live2.json"]);
+  test("script-diff-live", [jsonFile.path, jsonFile2.path]);
+
+  // This tests "dark-matter"-mode diffs.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-diff-dark-matter1.json"]);
+  jsonFile2 = FileUtils.getFile("CurWorkD", ["script-diff-dark-matter2.json"]);
+  test("script-diff-dark-matter", [jsonFile.path, jsonFile2.path]);
+}
diff --git a/memory/replace/dmd/test/xpcshell.ini b/memory/replace/dmd/test/xpcshell.ini
new file mode 100644
index 0000000000..7b4bdb2dad
--- /dev/null
+++ b/memory/replace/dmd/test/xpcshell.ini
@@ -0,0 +1,34 @@
+[DEFAULT]
+support-files =
+  basic-scan-32-expected.txt
+  basic-scan-64-expected.txt
+  complete-empty-live-expected.txt
+  complete-empty-dark-matter-expected.txt
+  complete-empty-cumulative-expected.txt
+  complete-full1-live-expected.txt
+  complete-full1-dark-matter-expected.txt
+  complete-full2-dark-matter-expected.txt
+  complete-full2-cumulative-expected.txt
+  complete-partial-live-expected.txt
+  scan-test.py
+  script-max-frames.json
+  script-max-frames-8-expected.txt
+  script-max-frames-3-expected.txt
+  script-max-frames-1-expected.txt
+  script-sort-by.json.gz
+  script-sort-by-usable-expected.txt
+  script-sort-by-req-expected.txt
+  script-sort-by-slop-expected.txt
+  script-sort-by-num-blocks-expected.txt
+  script-ignore-alloc-fns.json
+  script-ignore-alloc-fns-expected.txt
+  script-diff-live1.json
+  script-diff-live2.json
+  script-diff-live-expected.txt
+  script-diff-dark-matter1.json
+  script-diff-dark-matter2.json
+  script-diff-dark-matter-expected.txt
+
+[test_dmd.js]
+dmd = true
+skip-if = !(os=='linux' || os=='mac' || (os=='win' && !pgo)) || (os == "win" && processor == "aarch64") # aarch64 due to 1536250
diff --git a/memory/replace/logalloc/FdPrintf.cpp b/memory/replace/logalloc/FdPrintf.cpp
new file mode 100644
index 0000000000..d8e27ea133
--- /dev/null
+++ b/memory/replace/logalloc/FdPrintf.cpp
@@ -0,0 +1,199 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <cstdarg>
+
+#ifdef _WIN32
+#  include <windows.h>
+#else
+#  include <unistd.h>
+#endif
+#include <cmath>
+#include <cstring>
+#include "mozilla/Assertions.h"
+#include "mozilla/Unused.h"
+
+/* Template class allowing a limited number of increments on a value */
+template <typename T>
+class CheckedIncrement {
+ public:
+  CheckedIncrement(T aValue, size_t aMaxIncrement)
+      : mValue(aValue), mMaxIncrement(aMaxIncrement) {}
+
+  T operator++(int) {
+    if (!mMaxIncrement) {
+      MOZ_CRASH("overflow detected");
+    }
+    mMaxIncrement--;
+    return mValue++;
+  }
+
+  T& operator++() {
+    (*this)++;
+    return mValue;
+  }
+
+  void advance(T end) {
+    // Only makes sense if T is a pointer type.
+    size_t diff = end - mValue;
+    if (diff > mMaxIncrement) {
+      MOZ_CRASH("overflow detected");
+    }
+    mMaxIncrement -= diff;
+    mValue = end;
+  };
+
+  void rewind(T pos) {
+    size_t diff = mValue - pos;
+    mMaxIncrement += diff;
+    mValue = pos;
+  }
+
+  operator T() { return mValue; }
+  T value() { return mValue; }
+
+ private:
+  T mValue;
+  size_t mMaxIncrement;
+};
+
+template <typename T>
+static unsigned NumDigits(T n) {
+  if (n < 1) {
+    // We want one digit, it will be 0.
+    return 1;
+  }
+
+  double l = log10(static_cast<double>(n));
+  double cl = ceil(l);
+  return l == cl ? unsigned(cl) + 1 : unsigned(cl);
+}
+
+static void LeftPad(CheckedIncrement<char*>& b, size_t pad) {
+  while (pad-- > 0) {
+    *(b++) = ' ';
+  }
+}
+
+// Write the digits into the buffer.
+static void WriteDigits(CheckedIncrement<char*>& b, size_t i,
+                        size_t num_digits) {
+  size_t x = pow(10, double(num_digits - 1));
+  do {
+    *(b++) = "0123456789"[(i / x) % 10];
+    x /= 10;
+  } while (x > 0);
+}
+
+void FdPrintf(intptr_t aFd, const char* aFormat, ...) {
+  if (aFd == 0) {
+    return;
+  }
+  char buf[256];
+  CheckedIncrement<char*> b(buf, sizeof(buf));
+  CheckedIncrement<const char*> f(aFormat, strlen(aFormat) + 1);
+  va_list ap;
+  va_start(ap, aFormat);
+  while (true) {
+    switch (*f) {
+      case '\0':
+        goto out;
+
+      case '%': {
+        // The start of the format specifier is used if this specifier is
+        // invalid.
+        const char* start = f;
+
+        // Read the field width
+        f++;
+        char* end = nullptr;
+        size_t width = strtoul(f, &end, 10);
+        // If strtol can't find a number that's okay, that means 0 in our
+        // case, but we must advance f).
+        f.advance(end);
+
+        switch (*f) {
+          case 'z': {
+            if (*(++f) == 'u') {
+              size_t i = va_arg(ap, size_t);
+
+              size_t num_digits = NumDigits(i);
+              LeftPad(b, width > num_digits ? width - num_digits : 0);
+              WriteDigits(b, i, num_digits);
+            } else {
+              // If the format specifier is unknown then write out '%' and
+              // rewind to the beginning of the specifier causing it to be
+              // printed normally.
+              *(b++) = '%';
+              f.rewind(start);
+            }
+            break;
+          }
+
+          case 'p': {
+            intptr_t ptr = va_arg(ap, intptr_t);
+            *(b++) = '0';
+            *(b++) = 'x';
+            int x = sizeof(intptr_t) * 8;
+            bool wrote_msb = false;
+            do {
+              x -= 4;
+              size_t hex_digit = ptr >> x & 0xf;
+              if (hex_digit || wrote_msb) {
+                *(b++) = "0123456789abcdef"[hex_digit];
+                wrote_msb = true;
+              }
+            } while (x > 0);
+            if (!wrote_msb) {
+              *(b++) = '0';
+            }
+            break;
+          }
+
+          case 's': {
+            const char* str = va_arg(ap, const char*);
+            size_t len = strlen(str);
+
+            LeftPad(b, width > len ? width - len : 0);
+
+            while (*str) {
+              *(b++) = *(str++);
+            }
+
+            break;
+          }
+
+          case '%':
+            // Print a single raw '%'.
+            *(b++) = '%';
+            break;
+
+          default:
+            // If the format specifier is unknown then write out '%' and
+            // rewind to the beginning of the specifier causing it to be
+            // printed normally.
+            *(b++) = '%';
+            f.rewind(start);
+            break;
+        }
+        break;
+      }
+      default:
+        *(b++) = *f;
+        break;
+    }
+    f++;
+  }
+out:
+#ifdef _WIN32
+  // See comment in FdPrintf.h as to why WriteFile is used.
+  DWORD written;
+  WriteFile(reinterpret_cast<HANDLE>(aFd), buf, b - buf, &written, nullptr);
+#else
+  MOZ_UNUSED(write(aFd, buf, b - buf));
+#endif
+  va_end(ap);
+}
diff --git a/memory/replace/logalloc/FdPrintf.h b/memory/replace/logalloc/FdPrintf.h
new file mode 100644
index 0000000000..f390d57ed5
--- /dev/null
+++ b/memory/replace/logalloc/FdPrintf.h
@@ -0,0 +1,27 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef __FdPrintf_h__
+#define __FdPrintf_h__
+
+/* We can't use libc's (f)printf because it would reenter in replace_malloc,
+ * So use a custom and simplified version.  Only %p, %zu, %s and %% are
+ * supported, %zu, %s, support width specifiers.
+ *
+ * /!\ This function used a fixed-size internal buffer. The caller is
+ * expected to not use a format string that may overflow.
+ * The aFd argument is a file descriptor on UNIX and a native win32 file
+ * handle on Windows (from CreateFile). We can't use the windows POSIX
+ * APIs is that they don't support O_APPEND in a multi-process-safe way,
+ * while CreateFile does.
+ */
+extern void FdPrintf(intptr_t aFd, const char* aFormat, ...)
+#ifdef __GNUC__
+    __attribute__((format(printf, 2, 3)))
+#endif
+    ;
+
+#endif /* __FdPrintf_h__ */
diff --git a/memory/replace/logalloc/LogAlloc.cpp b/memory/replace/logalloc/LogAlloc.cpp
new file mode 100644
index 0000000000..8d6aad0675
--- /dev/null
+++ b/memory/replace/logalloc/LogAlloc.cpp
@@ -0,0 +1,236 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <cstdlib>
+#include <cstdio>
+#include <fcntl.h>
+
+#ifdef _WIN32
+#  include <windows.h>
+#  include <io.h>
+#  include <process.h>
+#else
+#  include <unistd.h>
+#  include <pthread.h>
+#endif
+
+#include "replace_malloc.h"
+#include "FdPrintf.h"
+#include "Mutex.h"
+
+static malloc_table_t sFuncs;
+static intptr_t sFd = 0;
+static bool sStdoutOrStderr = false;
+
+static Mutex sMutex;
+
+#ifndef _WIN32
+static void prefork() { sMutex.Lock(); }
+
+static void postfork() { sMutex.Unlock(); }
+#endif
+
+static size_t GetPid() { return size_t(getpid()); }
+
+static size_t GetTid() {
+#if defined(_WIN32)
+  return size_t(GetCurrentThreadId());
+#else
+  return size_t(pthread_self());
+#endif
+}
+
+#ifdef ANDROID
+/* Android doesn't have pthread_atfork defined in pthread.h */
+extern "C" MOZ_EXPORT int pthread_atfork(void (*)(void), void (*)(void),
+                                         void (*)(void));
+#endif
+
+class LogAllocBridge : public ReplaceMallocBridge {
+  virtual void InitDebugFd(mozilla::DebugFdRegistry& aRegistry) override {
+    if (!sStdoutOrStderr) {
+      aRegistry.RegisterHandle(sFd);
+    }
+  }
+};
+
+/* Do a simple, text-form, log of all calls to replace-malloc functions.
+ * Use locking to guarantee that an allocation that did happen is logged
+ * before any other allocation/free happens.
+ */
+
+static void* replace_malloc(size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.malloc(aSize);
+  FdPrintf(sFd, "%zu %zu malloc(%zu)=%p\n", GetPid(), GetTid(), aSize, ptr);
+  return ptr;
+}
+
+static int replace_posix_memalign(void** aPtr, size_t aAlignment,
+                                  size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  int ret = sFuncs.posix_memalign(aPtr, aAlignment, aSize);
+  FdPrintf(sFd, "%zu %zu posix_memalign(%zu,%zu)=%p\n", GetPid(), GetTid(),
+           aAlignment, aSize, (ret == 0) ? *aPtr : nullptr);
+  return ret;
+}
+
+static void* replace_aligned_alloc(size_t aAlignment, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.aligned_alloc(aAlignment, aSize);
+  FdPrintf(sFd, "%zu %zu aligned_alloc(%zu,%zu)=%p\n", GetPid(), GetTid(),
+           aAlignment, aSize, ptr);
+  return ptr;
+}
+
+static void* replace_calloc(size_t aNum, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.calloc(aNum, aSize);
+  FdPrintf(sFd, "%zu %zu calloc(%zu,%zu)=%p\n", GetPid(), GetTid(), aNum, aSize,
+           ptr);
+  return ptr;
+}
+
+static void* replace_realloc(void* aPtr, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* new_ptr = sFuncs.realloc(aPtr, aSize);
+  FdPrintf(sFd, "%zu %zu realloc(%p,%zu)=%p\n", GetPid(), GetTid(), aPtr, aSize,
+           new_ptr);
+  return new_ptr;
+}
+
+static void replace_free(void* aPtr) {
+  MutexAutoLock lock(sMutex);
+  FdPrintf(sFd, "%zu %zu free(%p)\n", GetPid(), GetTid(), aPtr);
+  sFuncs.free(aPtr);
+}
+
+static void* replace_memalign(size_t aAlignment, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.memalign(aAlignment, aSize);
+  FdPrintf(sFd, "%zu %zu memalign(%zu,%zu)=%p\n", GetPid(), GetTid(),
+           aAlignment, aSize, ptr);
+  return ptr;
+}
+
+static void* replace_valloc(size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.valloc(aSize);
+  FdPrintf(sFd, "%zu %zu valloc(%zu)=%p\n", GetPid(), GetTid(), aSize, ptr);
+  return ptr;
+}
+
+static void replace_jemalloc_stats(jemalloc_stats_t* aStats,
+                                   jemalloc_bin_stats_t* aBinStats) {
+  MutexAutoLock lock(sMutex);
+  sFuncs.jemalloc_stats_internal(aStats, aBinStats);
+  FdPrintf(sFd, "%zu %zu jemalloc_stats()\n", GetPid(), GetTid());
+}
+
+void replace_init(malloc_table_t* aTable, ReplaceMallocBridge** aBridge) {
+  /* Initialize output file descriptor from the MALLOC_LOG environment
+   * variable. Numbers up to 9999 are considered as a preopened file
+   * descriptor number. Other values are considered as a file name. */
+#ifdef _WIN32
+  wchar_t* log = _wgetenv(L"MALLOC_LOG");
+#else
+  char* log = getenv("MALLOC_LOG");
+#endif
+  if (log && *log) {
+    int fd = 0;
+    const auto* fd_num = log;
+    while (*fd_num) {
+      /* Reject non digits. */
+      if (*fd_num < '0' || *fd_num > '9') {
+        fd = -1;
+        break;
+      }
+      fd = fd * 10 + (*fd_num - '0');
+      /* Reject values >= 10000. */
+      if (fd >= 10000) {
+        fd = -1;
+        break;
+      }
+      fd_num++;
+    }
+    if (fd == 1 || fd == 2) {
+      sStdoutOrStderr = true;
+    }
+#ifdef _WIN32
+    // See comment in FdPrintf.h as to why CreateFile is used.
+    HANDLE handle;
+    if (fd > 0) {
+      handle = reinterpret_cast<HANDLE>(_get_osfhandle(fd));
+    } else {
+      handle =
+          CreateFileW(log, FILE_APPEND_DATA, FILE_SHARE_READ | FILE_SHARE_WRITE,
+                      nullptr, OPEN_ALWAYS, FILE_ATTRIBUTE_NORMAL, nullptr);
+    }
+    if (handle != INVALID_HANDLE_VALUE) {
+      sFd = reinterpret_cast<intptr_t>(handle);
+    }
+#else
+    if (fd == -1) {
+      fd = open(log, O_WRONLY | O_CREAT | O_APPEND, 0644);
+    }
+    if (fd > 0) {
+      sFd = fd;
+    }
+#endif
+  }
+
+  // Don't initialize if we weren't passed a valid MALLOC_LOG.
+  if (sFd == 0) {
+    return;
+  }
+
+  sMutex.Init();
+  static LogAllocBridge bridge;
+  sFuncs = *aTable;
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#define MALLOC_DECL(name, ...) aTable->name = replace_##name;
+#include "malloc_decls.h"
+  aTable->jemalloc_stats_internal = replace_jemalloc_stats;
+  if (!getenv("MALLOC_LOG_MINIMAL")) {
+    aTable->posix_memalign = replace_posix_memalign;
+    aTable->aligned_alloc = replace_aligned_alloc;
+    aTable->valloc = replace_valloc;
+  }
+  *aBridge = &bridge;
+
+#ifndef _WIN32
+  /* When another thread has acquired a lock before forking, the child
+   * process will inherit the lock state but the thread, being nonexistent
+   * in the child process, will never release it, leading to a dead-lock
+   * whenever the child process gets the lock. We thus need to ensure no
+   * other thread is holding the lock before forking, by acquiring it
+   * ourselves, and releasing it after forking, both in the parent and child
+   * processes.
+   * Windows doesn't have this problem since there is no fork().
+   * The real allocator, however, might be doing the same thing (jemalloc
+   * does). But pthread_atfork `prepare` handlers (first argument) are
+   * processed in reverse order they were established. But replace_init
+   * runs before the real allocator has had any chance to initialize and
+   * call pthread_atfork itself. This leads to its prefork running before
+   * ours. This leads to a race condition that can lead to a deadlock like
+   * the following:
+   *   - thread A forks.
+   *   - libc calls real allocator's prefork, so thread A holds the real
+   *     allocator lock.
+   *   - thread B calls malloc, which calls our replace_malloc.
+   *   - consequently, thread B holds our lock.
+   *   - thread B then proceeds to call the real allocator's malloc, and
+   *     waits for the real allocator's lock, which thread A holds.
+   *   - libc calls our prefork, so thread A waits for our lock, which
+   *     thread B holds.
+   * To avoid this race condition, the real allocator's prefork must be
+   * called after ours, which means it needs to be registered before ours.
+   * So trick the real allocator into initializing itself without more side
+   * effects by calling malloc with a size it can't possibly allocate. */
+  sFuncs.malloc(-1);
+  pthread_atfork(prefork, postfork, postfork);
+#endif
+}
diff --git a/memory/replace/logalloc/README b/memory/replace/logalloc/README
new file mode 100644
index 0000000000..c2e8cf66ce
--- /dev/null
+++ b/memory/replace/logalloc/README
@@ -0,0 +1,95 @@
+Logalloc is a replace-malloc library for Firefox (see
+memory/build/replace_malloc.h) that dumps a log of memory allocations to a
+given file descriptor or file name. That log can then be replayed against
+Firefox's default memory allocator independently or through another
+replace-malloc library, allowing the testing of other allocators under the
+exact same workload.
+
+To get an allocation log the following environment variable when starting
+Firefox:
+  MALLOC_LOG=/path/to/log-file
+  or
+  MALLOC_LOG=number
+
+When MALLOC_LOG is a number below 10000, it is considered as a file
+descriptor number that is fed to Firefox when it is started. Otherwise,
+it is considered as a file name.
+
+As those allocation logs can grow large quite quickly, it can be useful
+to pipe the output to a compression tool.
+
+MALLOC_LOG=1 would send to Firefox's stdout, MALLOC_LOG=2 would send to
+its stderr. Since in both cases that could be mixed with other output
+from Firefox, it is usually better to use another file descriptor
+by shell redirections, such as:
+
+  MALLOC_LOG=3 firefox 3>&1 1>&2 | gzip -c > log.gz
+
+(3>&1 copies the `| gzip` pipe file descriptor to file descriptor #3, 1>&2
+then copies stderr to stdout. This leads to: fd1 and fd2 sending to stderr
+of the parent process (the shell), and fd3 sending to gzip.)
+
+Each line of the allocations log is formatted as follows:
+  <pid> <tid> <function>([<args>])[=<result>]
+where <args> is a comma separated list of values. The number of <args> and
+the presence of <result> depend on the <function>.
+
+Example log:
+  18545 18545 malloc(32)=0x7f90495120e0
+  18545 18545 calloc(1,148)=0x7f9049537480
+  18545 18545 realloc(0x7f90495120e0,64)=0x7f9049536680
+  18545 18545 posix_memalign(256,240)=0x7f9049583300
+  18545 18545 jemalloc_stats()
+  18545 18545 free(0x7f9049536680)
+
+This log can be replayed with the logalloc-replay tool in
+memory/replace/logalloc/replay. However, as the goal of that tool is to
+reproduce the recorded memory allocations, it needs to avoid as much as
+possible doing its own allocations for bookkeeping. Reading the logs as
+they are would require data structures and memory allocations. As a
+consequence, the logs need to be preprocessed beforehand.
+
+The logalloc_munge.py script is responsible for that preprocessing. It simply
+takes a raw log on its stdin, and outputs the preprocessed log on its stdout.
+It replaces pointer addresses with indexes the logalloc-replay tool can use
+in a large (almost) linear array of allocation tracking slots (prefixed with
+'#'). It also replaces the pids with numbers starting from 1 (such as the
+first seen pid number is 1, the second is 2, etc.).
+
+The above example log would become the following, once preprocessed:
+  1 1 malloc(32)=#1
+  1 1 calloc(1,148)=#2
+  1 1 realloc(#1,64)=#1
+  1 1 posix_memalign(256,240)=#3
+  1 1 jemalloc_stats()
+  1 1 free(#1)
+
+The logalloc-replay tool then takes the preprocessed log on its stdin and
+replays the allocations printed there, but will only replay those with the
+same process id as the first line (which normally is 1).
+
+As the log files are simple text files, though, it is easy to separate out
+the different processes log with e.g. grep, and feed the separate processes
+logs to logalloc-replay.
+
+The logalloc-replay program won't output anything unless jemalloc_stats
+records appears in the log. You can expect those to be recorded when going
+to about:memory in Firefox, but they can also be added after preprocessing.
+
+Here is an example of what one can do:
+
+  gunzip -c log.gz | python logalloc_munge.py | \
+  awk '$1 == "2" { print $0 } !(NR % 10000) { print "2 1 jemalloc_stats()" }' | \
+    ./logalloc-replay
+
+The above command replays the allocations of process #2, with some stats
+output every 10000 records.
+
+The logalloc-replay tool itself being hooked with replace-malloc, it is possible
+to set LD_PRELOAD/DYLD_INSERT_LIBRARIES/MOZ_REPLACE_MALLOC_LIB and replay a log
+through a different allocator. For example:
+
+  LD_PRELOAD=libreplace_jemalloc.so logalloc-replay < log
+
+Will replay the log against jemalloc4 (which is, as of writing, what
+libreplace_jemalloc.so contains).
diff --git a/memory/replace/logalloc/moz.build b/memory/replace/logalloc/moz.build
new file mode 100644
index 0000000000..c52d9e69e0
--- /dev/null
+++ b/memory/replace/logalloc/moz.build
@@ -0,0 +1,30 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+ReplaceMalloc("logalloc")
+
+SOURCES += [
+    "FdPrintf.cpp",
+    "LogAlloc.cpp",
+]
+
+DisableStlWrapping()
+NO_PGO = True
+DEFINES["MOZ_NO_MOZALLOC"] = True
+
+LOCAL_INCLUDES += [
+    "/memory/build",
+]
+
+# Android doesn't have pthread_atfork, but we have our own in mozglue.
+if CONFIG["OS_TARGET"] == "Android" and FORCE_SHARED_LIB:
+    USE_LIBS += [
+        "mozglue",
+    ]
+
+DIRS += [
+    "replay",
+]
diff --git a/memory/replace/logalloc/replay/Makefile.in b/memory/replace/logalloc/replay/Makefile.in
new file mode 100644
index 0000000000..8f5b5589b6
--- /dev/null
+++ b/memory/replace/logalloc/replay/Makefile.in
@@ -0,0 +1,42 @@
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+ifndef CROSS_COMPILE
+ifndef MOZ_CODE_COVERAGE
+
+ifeq ($(OS_TARGET),WINNT)
+LOGALLOC_VAR = MOZ_REPLACE_MALLOC_LIB
+else
+ifeq ($(OS_TARGET),Darwin)
+LOGALLOC_VAR = DYLD_INSERT_LIBRARIES
+else
+LOGALLOC_VAR = LD_PRELOAD
+endif
+endif
+
+ifndef MOZ_REPLACE_MALLOC_STATIC
+LOGALLOC = $(LOGALLOC_VAR)=$(CURDIR)/../$(DLL_PREFIX)logalloc$(DLL_SUFFIX)
+endif
+
+expected_output.log: $(srcdir)/replay.log
+# The logalloc-replay program will only replay entries from the first pid,
+# so the expected output only contains entries beginning with "1 "
+	grep "^1 " $< > $@
+
+check:: $(srcdir)/replay.log expected_output.log $(srcdir)/expected_output_minimal.log
+# Test with MALLOC_LOG as a file descriptor number
+# We filter out anything happening before the first jemalloc_stats (first
+# command in replay.log) because starting with libstdc++ 5, a static
+# initializer in the STL allocates memory, which we obviously don't have
+# in expected_output.log.
+	MALLOC_LOG=1 $(LOGALLOC) ./$(PROGRAM) < $< | sed -n '/jemalloc_stats/,$$p' | $(PYTHON3) $(srcdir)/logalloc_munge.py | diff -w - expected_output.log
+# Test with MALLOC_LOG as a file name
+	$(RM) test_output.log
+	MALLOC_LOG=test_output.log $(LOGALLOC) ./$(PROGRAM) < $<
+	sed -n '/jemalloc_stats/,$$p' test_output.log | $(PYTHON3) $(srcdir)/logalloc_munge.py | diff -w - expected_output.log
+
+	MALLOC_LOG=1 MALLOC_LOG_MINIMAL=1 $(LOGALLOC) ./$(PROGRAM) < $< | sed -n '/jemalloc_stats/,$$p' | $(PYTHON3) $(srcdir)/logalloc_munge.py | diff -w - $(srcdir)/expected_output_minimal.log
+
+endif
+endif
diff --git a/memory/replace/logalloc/replay/Replay.cpp b/memory/replace/logalloc/replay/Replay.cpp
new file mode 100644
index 0000000000..daf0580389
--- /dev/null
+++ b/memory/replace/logalloc/replay/Replay.cpp
@@ -0,0 +1,739 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#define MOZ_MEMORY_IMPL
+#include "mozmemory_wrap.h"
+
+#ifdef _WIN32
+#  include <windows.h>
+#  include <io.h>
+typedef intptr_t ssize_t;
+#else
+#  include <sys/mman.h>
+#  include <unistd.h>
+#endif
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/MathAlgorithms.h"
+#include "FdPrintf.h"
+
+static void die(const char* message) {
+  /* Here, it doesn't matter that fprintf may allocate memory. */
+  fprintf(stderr, "%s\n", message);
+  exit(1);
+}
+
+/* We don't want to be using malloc() to allocate our internal tracking
+ * data, because that would change the parameters of what is being measured,
+ * so we want to use data types that directly use mmap/VirtualAlloc. */
+template <typename T, size_t Len>
+class MappedArray {
+ public:
+  MappedArray() : mPtr(nullptr) {}
+
+  ~MappedArray() {
+    if (mPtr) {
+#ifdef _WIN32
+      VirtualFree(mPtr, sizeof(T) * Len, MEM_RELEASE);
+#else
+      munmap(mPtr, sizeof(T) * Len);
+#endif
+    }
+  }
+
+  T& operator[](size_t aIndex) const {
+    if (mPtr) {
+      return mPtr[aIndex];
+    }
+
+#ifdef _WIN32
+    mPtr = reinterpret_cast<T*>(VirtualAlloc(
+        nullptr, sizeof(T) * Len, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE));
+    if (mPtr == nullptr) {
+      die("VirtualAlloc error");
+    }
+#else
+    mPtr = reinterpret_cast<T*>(mmap(nullptr, sizeof(T) * Len,
+                                     PROT_READ | PROT_WRITE,
+                                     MAP_ANON | MAP_PRIVATE, -1, 0));
+    if (mPtr == MAP_FAILED) {
+      die("Mmap error");
+    }
+#endif
+    return mPtr[aIndex];
+  }
+
+ private:
+  mutable T* mPtr;
+};
+
+/* Type for records of allocations. */
+struct MemSlot {
+  void* mPtr;
+
+  // mRequest is only valid if mPtr is non-null.  It doesn't need to be cleared
+  // when memory is freed or realloc()ed.
+  size_t mRequest;
+};
+
+/* An almost infinite list of slots.
+ * In essence, this is a linked list of arrays of groups of slots.
+ * Each group is 1MB. On 64-bits, one group allows to store 64k allocations.
+ * Each MemSlotList instance can store 1023 such groups, which means more
+ * than 67M allocations. In case more would be needed, we chain to another
+ * MemSlotList, and so on.
+ * Using 1023 groups makes the MemSlotList itself page sized on 32-bits
+ * and 2 pages-sized on 64-bits.
+ */
+class MemSlotList {
+  static const size_t kGroups = 1024 - 1;
+  static const size_t kGroupSize = (1024 * 1024) / sizeof(MemSlot);
+
+  MappedArray<MemSlot, kGroupSize> mSlots[kGroups];
+  MappedArray<MemSlotList, 1> mNext;
+
+ public:
+  MemSlot& operator[](size_t aIndex) const {
+    if (aIndex < kGroupSize * kGroups) {
+      return mSlots[aIndex / kGroupSize][aIndex % kGroupSize];
+    }
+    aIndex -= kGroupSize * kGroups;
+    return mNext[0][aIndex];
+  }
+};
+
+/* Helper class for memory buffers */
+class Buffer {
+ public:
+  Buffer() : mBuf(nullptr), mLength(0) {}
+
+  Buffer(const void* aBuf, size_t aLength)
+      : mBuf(reinterpret_cast<const char*>(aBuf)), mLength(aLength) {}
+
+  /* Constructor for string literals. */
+  template <size_t Size>
+  explicit Buffer(const char (&aStr)[Size]) : mBuf(aStr), mLength(Size - 1) {}
+
+  /* Returns a sub-buffer up-to but not including the given aNeedle character.
+   * The "parent" buffer itself is altered to begin after the aNeedle
+   * character.
+   * If the aNeedle character is not found, return the entire buffer, and empty
+   * the "parent" buffer. */
+  Buffer SplitChar(char aNeedle) {
+    char* buf = const_cast<char*>(mBuf);
+    char* c = reinterpret_cast<char*>(memchr(buf, aNeedle, mLength));
+    if (!c) {
+      return Split(mLength);
+    }
+
+    Buffer result = Split(c - buf);
+    // Remove the aNeedle character itself.
+    Split(1);
+    return result;
+  }
+
+  /* Returns a sub-buffer of at most aLength characters. The "parent" buffer is
+   * amputated of those aLength characters. If the "parent" buffer is smaller
+   * than aLength, then its length is used instead. */
+  Buffer Split(size_t aLength) {
+    Buffer result(mBuf, std::min(aLength, mLength));
+    mLength -= result.mLength;
+    mBuf += result.mLength;
+    return result;
+  }
+
+  /* Move the buffer (including its content) to the memory address of the aOther
+   * buffer. */
+  void Slide(Buffer aOther) {
+    memmove(const_cast<char*>(aOther.mBuf), mBuf, mLength);
+    mBuf = aOther.mBuf;
+  }
+
+  /* Returns whether the two involved buffers have the same content. */
+  bool operator==(Buffer aOther) {
+    return mLength == aOther.mLength &&
+           (mBuf == aOther.mBuf || !strncmp(mBuf, aOther.mBuf, mLength));
+  }
+
+  /* Returns whether the buffer is empty. */
+  explicit operator bool() { return mLength; }
+
+  /* Returns the memory location of the buffer. */
+  const char* get() { return mBuf; }
+
+  /* Returns the memory location of the end of the buffer (technically, the
+   * first byte after the buffer). */
+  const char* GetEnd() { return mBuf + mLength; }
+
+  /* Extend the buffer over the content of the other buffer, assuming it is
+   * adjacent. */
+  void Extend(Buffer aOther) {
+    MOZ_ASSERT(aOther.mBuf == GetEnd());
+    mLength += aOther.mLength;
+  }
+
+ private:
+  const char* mBuf;
+  size_t mLength;
+};
+
+/* Helper class to read from a file descriptor line by line. */
+class FdReader {
+ public:
+  explicit FdReader(int aFd)
+      : mFd(aFd), mData(&mRawBuf, 0), mBuf(&mRawBuf, sizeof(mRawBuf)) {}
+
+  /* Read a line from the file descriptor and returns it as a Buffer instance */
+  Buffer ReadLine() {
+    while (true) {
+      Buffer result = mData.SplitChar('\n');
+
+      /* There are essentially three different cases here:
+       * - '\n' was found "early". In this case, the end of the result buffer
+       *   is before the beginning of the mData buffer (since SplitChar
+       *   amputated it).
+       * - '\n' was found as the last character of mData. In this case, mData
+       *   is empty, but still points at the end of mBuf. result points to what
+       *   used to be in mData, without the last character.
+       * - '\n' was not found. In this case too, mData is empty and points at
+       *   the end of mBuf. But result points to the entire buffer that used to
+       *   be pointed by mData.
+       * Only in the latter case do both result and mData's end match, and it's
+       * the only case where we need to refill the buffer.
+       */
+      if (result.GetEnd() != mData.GetEnd()) {
+        return result;
+      }
+
+      /* Since SplitChar emptied mData, make it point to what it had before. */
+      mData = result;
+
+      /* And move it to the beginning of the read buffer. */
+      mData.Slide(mBuf);
+
+      FillBuffer();
+
+      if (!mData) {
+        return Buffer();
+      }
+    }
+  }
+
+ private:
+  /* Fill the read buffer. */
+  void FillBuffer() {
+    size_t size = mBuf.GetEnd() - mData.GetEnd();
+    Buffer remainder(mData.GetEnd(), size);
+
+    ssize_t len = 1;
+    while (remainder && len > 0) {
+      len = ::read(mFd, const_cast<char*>(remainder.get()), size);
+      if (len < 0) {
+        die("Read error");
+      }
+      size -= len;
+      mData.Extend(remainder.Split(len));
+    }
+  }
+
+  /* File descriptor to read from. */
+  int mFd;
+  /* Part of data that was read from the file descriptor but not returned with
+   * ReadLine yet. */
+  Buffer mData;
+  /* Buffer representation of mRawBuf */
+  Buffer mBuf;
+  /* read() buffer */
+  char mRawBuf[4096];
+};
+
+MOZ_BEGIN_EXTERN_C
+
+/* Function declarations for all the replace_malloc _impl functions.
+ * See memory/build/replace_malloc.c */
+#define MALLOC_DECL(name, return_type, ...) \
+  return_type name##_impl(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC
+#include "malloc_decls.h"
+
+#define MALLOC_DECL(name, return_type, ...) return_type name(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_JEMALLOC
+#include "malloc_decls.h"
+
+#ifdef ANDROID
+
+/* mozjemalloc and jemalloc use pthread_atfork, which Android doesn't have.
+ * While gecko has one in libmozglue, the replay program can't use that.
+ * Since we're not going to fork anyways, make it a dummy function. */
+int pthread_atfork(void (*aPrepare)(void), void (*aParent)(void),
+                   void (*aChild)(void)) {
+  return 0;
+}
+#endif
+
+MOZ_END_EXTERN_C
+
+size_t parseNumber(Buffer aBuf) {
+  if (!aBuf) {
+    die("Malformed input");
+  }
+
+  size_t result = 0;
+  for (const char *c = aBuf.get(), *end = aBuf.GetEnd(); c < end; c++) {
+    if (*c < '0' || *c > '9') {
+      die("Malformed input");
+    }
+    result *= 10;
+    result += *c - '0';
+  }
+  return result;
+}
+
+/* Class to handle dispatching the replay function calls to replace-malloc. */
+class Replay {
+ public:
+  Replay()
+      : mOps(0),
+        mNumUsedSlots(0),
+        mTotalRequestedSize(0),
+        mTotalAllocatedSize(0),
+        mCalculateSlop(false) {
+#ifdef _WIN32
+    // See comment in FdPrintf.h as to why native win32 handles are used.
+    mStdErr = reinterpret_cast<intptr_t>(GetStdHandle(STD_ERROR_HANDLE));
+#else
+    mStdErr = fileno(stderr);
+#endif
+  }
+
+  void enableSlopCalculation() { mCalculateSlop = true; }
+
+  MemSlot& operator[](size_t index) const { return mSlots[index]; }
+
+  void malloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::malloc_impl(size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void posix_memalign(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t alignment = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    void* ptr;
+    if (::posix_memalign_impl(&ptr, alignment, size) == 0) {
+      aSlot.mPtr = ptr;
+      aSlot.mRequest = size;
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    } else {
+      aSlot.mPtr = nullptr;
+    }
+  }
+
+  void aligned_alloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t alignment = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::aligned_alloc_impl(alignment, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void calloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t num = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::calloc_impl(num, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = num * size;
+      if (mCalculateSlop) {
+        mTotalRequestedSize += num * size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void realloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    Buffer dummy = aArgs.SplitChar('#');
+    if (dummy) {
+      die("Malformed input");
+    }
+    size_t slot_id = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    MemSlot& old_slot = (*this)[slot_id];
+    void* old_ptr = old_slot.mPtr;
+    old_slot.mPtr = nullptr;
+    aSlot.mPtr = ::realloc_impl(old_ptr, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void free(Buffer& aArgs, Buffer& aResult) {
+    if (aResult) {
+      die("Malformed input");
+    }
+    mOps++;
+    Buffer dummy = aArgs.SplitChar('#');
+    if (dummy) {
+      die("Malformed input");
+    }
+    size_t slot_id = parseNumber(aArgs);
+    MemSlot& slot = (*this)[slot_id];
+    ::free_impl(slot.mPtr);
+    slot.mPtr = nullptr;
+  }
+
+  void memalign(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t alignment = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::memalign_impl(alignment, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void valloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::valloc_impl(size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void jemalloc_stats(Buffer& aArgs, Buffer& aResult) {
+    if (aArgs || aResult) {
+      die("Malformed input");
+    }
+    mOps++;
+    jemalloc_stats_t stats;
+    jemalloc_bin_stats_t bin_stats[JEMALLOC_MAX_STATS_BINS];
+    ::jemalloc_stats_internal(&stats, bin_stats);
+
+    size_t num_objects = 0;
+    size_t num_sloppy_objects = 0;
+    size_t total_allocated = 0;
+    size_t total_slop = 0;
+    size_t large_slop = 0;
+    size_t large_used = 0;
+    size_t huge_slop = 0;
+    size_t huge_used = 0;
+    size_t bin_slop[JEMALLOC_MAX_STATS_BINS] = {0};
+
+    for (size_t slot_id = 0; slot_id < mNumUsedSlots; slot_id++) {
+      MemSlot& slot = mSlots[slot_id];
+      if (slot.mPtr) {
+        size_t used = ::malloc_usable_size_impl(slot.mPtr);
+        size_t slop = used - slot.mRequest;
+        total_allocated += used;
+        total_slop += slop;
+        num_objects++;
+        if (slop) {
+          num_sloppy_objects++;
+        }
+
+        if (used <= stats.page_size / 2) {
+          // We know that this is an inefficient linear search, but there's a
+          // small number of bins and this is simple.
+          for (unsigned i = 0; i < JEMALLOC_MAX_STATS_BINS; i++) {
+            auto& bin = bin_stats[i];
+            if (used == bin.size) {
+              bin_slop[i] += slop;
+              break;
+            }
+          }
+        } else if (used <= stats.large_max) {
+          large_slop += slop;
+          large_used += used;
+        } else {
+          huge_slop += slop;
+          huge_used += used;
+        }
+      }
+    }
+
+    FdPrintf(mStdErr, "\n");
+    FdPrintf(mStdErr, "Objects:      %9zu\n", num_objects);
+    FdPrintf(mStdErr, "Slots:        %9zu\n", mNumUsedSlots);
+    FdPrintf(mStdErr, "Ops:          %9zu\n", mOps);
+    FdPrintf(mStdErr, "mapped:       %9zu\n", stats.mapped);
+    FdPrintf(mStdErr, "allocated:    %9zu\n", stats.allocated);
+    FdPrintf(mStdErr, "waste:        %9zu\n", stats.waste);
+    FdPrintf(mStdErr, "dirty:        %9zu\n", stats.page_cache);
+    FdPrintf(mStdErr, "bookkeep:     %9zu\n", stats.bookkeeping);
+    FdPrintf(mStdErr, "bin-unused:   %9zu\n", stats.bin_unused);
+    FdPrintf(mStdErr, "quantum-max:  %9zu\n", stats.quantum_max);
+    FdPrintf(mStdErr, "subpage-max:  %9zu\n", stats.page_size / 2);
+    FdPrintf(mStdErr, "large-max:    %9zu\n", stats.large_max);
+    if (mCalculateSlop) {
+      size_t slop = mTotalAllocatedSize - mTotalRequestedSize;
+      FdPrintf(mStdErr,
+               "Total slop for all allocations: %zuKiB/%zuKiB (%zu%%)\n",
+               slop / 1024, mTotalAllocatedSize / 1024,
+               percent(slop, mTotalAllocatedSize));
+    }
+    FdPrintf(mStdErr, "Live sloppy objects: %zu/%zu (%zu%%)\n",
+             num_sloppy_objects, num_objects,
+             percent(num_sloppy_objects, num_objects));
+    FdPrintf(mStdErr, "Live sloppy bytes: %zuKiB/%zuKiB (%zu%%)\n",
+             total_slop / 1024, total_allocated / 1024,
+             percent(total_slop, total_allocated));
+
+    FdPrintf(mStdErr, "\n%8s %11s %10s %8s %9s %9s %8s\n", "bin-size",
+             "unused (c)", "total (c)", "used (c)", "non-full (r)", "total (r)",
+             "used (r)");
+    for (auto& bin : bin_stats) {
+      if (bin.size) {
+        FdPrintf(mStdErr, "%8zu %8zuKiB %7zuKiB %7zu%% %12zu %9zu %7zu%%\n",
+                 bin.size, bin.bytes_unused / 1024, bin.bytes_total / 1024,
+                 percent(bin.bytes_total - bin.bytes_unused, bin.bytes_total),
+                 bin.num_non_full_runs, bin.num_runs,
+                 percent(bin.num_runs - bin.num_non_full_runs, bin.num_runs));
+      }
+    }
+
+    FdPrintf(mStdErr, "\n%5s %8s %9s %7s\n", "bin", "slop", "used", "percent");
+    for (unsigned i = 0; i < JEMALLOC_MAX_STATS_BINS; i++) {
+      auto& bin = bin_stats[i];
+      if (bin.size) {
+        size_t used = bin.bytes_total - bin.bytes_unused;
+        FdPrintf(mStdErr, "%5zu %8zu %9zu %6zu%%\n", bin.size, bin_slop[i],
+                 used, percent(bin_slop[i], used));
+      }
+    }
+    FdPrintf(mStdErr, "%5s %8zu %9zu %6zu%%\n", "large", large_slop, large_used,
+             percent(large_slop, large_used));
+    FdPrintf(mStdErr, "%5s %8zu %9zu %6zu%%\n", "huge", huge_slop, huge_used,
+             percent(huge_slop, huge_used));
+
+    unsigned last_size = 0;
+    for (auto& bin : bin_stats) {
+      if (bin.size == 0) {
+        break;
+      }
+
+      if (bin.size <= 16) {
+        // 1 byte buckets.
+        print_distribution(bin.size, last_size, 1);
+      } else if (bin.size <= stats.quantum_max) {
+        // 4 buckets, (4 bytes per bucket with a 16 byte quantum).
+        print_distribution(bin.size, last_size, stats.quantum / 4);
+      } else {
+        // 16 buckets.
+        print_distribution(bin.size, last_size, (bin.size - last_size) / 16);
+      }
+
+      last_size = bin.size;
+    }
+
+    // 16 buckets.
+    print_distribution(stats.page_size, last_size,
+                       (stats.page_size - last_size) / 16);
+
+    // Buckets are 1/4 of the page size (12 buckets).
+    print_distribution(stats.page_size * 4, stats.page_size,
+                       stats.page_size / 4);
+
+    /* TODO: Add more data, like actual RSS as measured by OS, but compensated
+     * for the replay internal data. */
+  }
+
+ private:
+  const size_t MAX_NUM_BUCKETS = 16;
+
+  /*
+   * Create and print frequency distributions of memory requests.
+   */
+  void print_distribution(size_t size, size_t next_smallest,
+                          size_t bucket_size) {
+    unsigned shift = mozilla::CeilingLog2(bucket_size);
+
+    // The number of slots.
+    const unsigned array_slots = (size - next_smallest) >> shift;
+
+    // The translation to turn a slot index into a memory request size.
+    const unsigned array_offset = 1 + next_smallest;
+    const size_t array_offset_add = (1 << shift) + next_smallest;
+
+    // Avoid a variable length array.
+    MOZ_RELEASE_ASSERT(array_slots <= MAX_NUM_BUCKETS);
+    size_t requests[MAX_NUM_BUCKETS];
+    memset(requests, 0, sizeof(size_t) * array_slots);
+    size_t total_requests = 0;
+
+    for (size_t slot_id = 0; slot_id < mNumUsedSlots; slot_id++) {
+      MemSlot& slot = mSlots[slot_id];
+      if (slot.mPtr && slot.mRequest > next_smallest && slot.mRequest <= size) {
+        requests[(slot.mRequest - array_offset) >> shift]++;
+        total_requests++;
+      }
+    }
+
+    FdPrintf(mStdErr, "\n%zu-bin Distribution:\n", size);
+    FdPrintf(mStdErr, "   request   :  count percent\n");
+    size_t range_start = next_smallest + 1;
+    for (size_t j = 0; j < array_slots; j++) {
+      size_t range_end = (j << shift) + array_offset_add;
+      FdPrintf(mStdErr, "%5zu - %5zu: %6zu %6zu%%\n", range_start, range_end,
+               requests[j], percent(requests[j], total_requests));
+      range_start = range_end + 1;
+    }
+  }
+
+  static size_t percent(size_t a, size_t b) {
+    if (!b) {
+      return 0;
+    }
+    return size_t(round(double(a) / double(b) * 100.0));
+  }
+
+  MemSlot& SlotForResult(Buffer& aResult) {
+    /* Parse result value and get the corresponding slot. */
+    Buffer dummy = aResult.SplitChar('=');
+    Buffer dummy2 = aResult.SplitChar('#');
+    if (dummy || dummy2) {
+      die("Malformed input");
+    }
+
+    size_t slot_id = parseNumber(aResult);
+    mNumUsedSlots = std::max(mNumUsedSlots, slot_id + 1);
+
+    return mSlots[slot_id];
+  }
+
+  intptr_t mStdErr;
+  size_t mOps;
+
+  // The number of slots that have been used. It is used to iterate over slots
+  // without accessing those we haven't initialised.
+  size_t mNumUsedSlots;
+
+  MemSlotList mSlots;
+  size_t mTotalRequestedSize;
+  size_t mTotalAllocatedSize;
+  // Whether to calculate slop for all allocations over the runtime of a
+  // process.
+  bool mCalculateSlop;
+};
+
+int main(int argc, const char* argv[]) {
+  size_t first_pid = 0;
+  FdReader reader(0);
+  Replay replay;
+
+  for (int i = 1; i < argc; i++) {
+    const char* option = argv[i];
+    if (strcmp(option, "-s") == 0) {
+      replay.enableSlopCalculation();
+    } else {
+      fprintf(stderr, "Unknown command line option: %s\n", option);
+      return EXIT_FAILURE;
+    }
+  }
+
+  /* Read log from stdin and dispatch function calls to the Replay instance.
+   * The log format is essentially:
+   *   <pid> <tid> <function>([<args>])[=<result>]
+   * <args> is a comma separated list of arguments.
+   *
+   * The logs are expected to be preprocessed so that allocations are
+   * attributed a tracking slot. The input is trusted not to have crazy
+   * values for these slot numbers.
+   *
+   * <result>, as well as some of the args to some of the function calls are
+   * such slot numbers.
+   */
+  while (true) {
+    Buffer line = reader.ReadLine();
+
+    if (!line) {
+      break;
+    }
+
+    size_t pid = parseNumber(line.SplitChar(' '));
+    if (!first_pid) {
+      first_pid = pid;
+    }
+
+    /* The log may contain data for several processes, only entries for the
+     * very first that appears are treated. */
+    if (first_pid != pid) {
+      continue;
+    }
+
+    /* The log contains thread ids for manual analysis, but we just ignore them
+     * for now. */
+    parseNumber(line.SplitChar(' '));
+
+    Buffer func = line.SplitChar('(');
+    Buffer args = line.SplitChar(')');
+
+    if (func == Buffer("jemalloc_stats")) {
+      replay.jemalloc_stats(args, line);
+    } else if (func == Buffer("free")) {
+      replay.free(args, line);
+    } else if (func == Buffer("malloc")) {
+      replay.malloc(args, line);
+    } else if (func == Buffer("posix_memalign")) {
+      replay.posix_memalign(args, line);
+    } else if (func == Buffer("aligned_alloc")) {
+      replay.aligned_alloc(args, line);
+    } else if (func == Buffer("calloc")) {
+      replay.calloc(args, line);
+    } else if (func == Buffer("realloc")) {
+      replay.realloc(args, line);
+    } else if (func == Buffer("memalign")) {
+      replay.memalign(args, line);
+    } else if (func == Buffer("valloc")) {
+      replay.valloc(args, line);
+    } else {
+      die("Malformed input");
+    }
+  }
+
+  return 0;
+}
diff --git a/memory/replace/logalloc/replay/expected_output_minimal.log b/memory/replace/logalloc/replay/expected_output_minimal.log
new file mode 100644
index 0000000000..332fe20957
--- /dev/null
+++ b/memory/replace/logalloc/replay/expected_output_minimal.log
@@ -0,0 +1,17 @@
+1 1 jemalloc_stats()
+1 1 malloc(42)=#1
+1 1 malloc(24)=#2
+1 1 free(#1)
+1 1 memalign(4096,1024)=#1
+1 1 calloc(4,42)=#3
+1 1 free(#2)
+1 1 realloc(#3,84)=#2
+1 1 memalign(256,1024)=#3
+1 1 memalign(512,1024)=#4
+1 1 memalign(4096,1024)=#5
+1 1 jemalloc_stats()
+1 1 free(#5)
+1 1 free(#4)
+1 1 free(#3)
+1 1 free(#2)
+1 1 free(#1)
diff --git a/memory/replace/logalloc/replay/logalloc_munge.py b/memory/replace/logalloc/replay/logalloc_munge.py
new file mode 100644
index 0000000000..bea397d505
--- /dev/null
+++ b/memory/replace/logalloc/replay/logalloc_munge.py
@@ -0,0 +1,151 @@
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+"""
+This script takes a log from the replace-malloc logalloc library on stdin
+and munges it so that it can be used with the logalloc-replay tool.
+
+Given the following output:
+  13663 malloc(42)=0x7f0c33502040
+  13663 malloc(24)=0x7f0c33503040
+  13663 free(0x7f0c33502040)
+The resulting output is:
+  1 malloc(42)=#1
+  1 malloc(24)=#2
+  1 free(#1)
+
+See README for more details.
+"""
+
+from __future__ import absolute_import, print_function
+import sys
+from collections import (
+    defaultdict,
+    deque,
+)
+
+
+class IdMapping(object):
+    """Class to map values to ids.
+
+    Each value is associated to an increasing id, starting from 1.
+    When a value is removed, its id is recycled and will be reused for
+    subsequent values.
+    """
+
+    def __init__(self):
+        self.id = 1
+        self._values = {}
+        self._recycle = deque()
+
+    def __getitem__(self, value):
+        if value not in self._values:
+            if self._recycle:
+                self._values[value] = self._recycle.popleft()
+            else:
+                self._values[value] = self.id
+                self.id += 1
+        return self._values[value]
+
+    def __delitem__(self, value):
+        if value == 0:
+            return
+        self._recycle.append(self._values[value])
+        del self._values[value]
+
+    def __contains__(self, value):
+        return value == 0 or value in self._values
+
+
+class Ignored(Exception):
+    pass
+
+
+def split_log_line(line):
+    try:
+        # The format for each line is:
+        # <pid> [<tid>] <function>([<args>])[=<result>]
+        #
+        # The original format didn't include the tid, so we try to parse
+        # lines whether they have one or not.
+        pid, func_call = line.split(" ", 1)
+        call, result = func_call.split(")")
+        func, args = call.split("(")
+        args = args.split(",") if args else []
+        if result:
+            if result[0] != "=":
+                raise Ignored("Malformed input")
+            result = result[1:]
+        if " " in func:
+            tid, func = func.split(" ", 1)
+        else:
+            tid = pid
+        return pid, tid, func, args, result
+    except Exception:
+        raise Ignored("Malformed input")
+
+
+NUM_ARGUMENTS = {
+    "jemalloc_stats": 0,
+    "free": 1,
+    "malloc": 1,
+    "posix_memalign": 2,
+    "aligned_alloc": 2,
+    "calloc": 2,
+    "realloc": 2,
+    "memalign": 2,
+    "valloc": 1,
+}
+
+
+def main():
+    pids = IdMapping()
+    processes = defaultdict(lambda: {"pointers": IdMapping(), "tids": IdMapping()})
+    for line in sys.stdin:
+        line = line.strip()
+
+        try:
+            pid, tid, func, args, result = split_log_line(line)
+
+            # Replace pid with an id.
+            pid = pids[int(pid)]
+
+            process = processes[pid]
+            tid = process["tids"][int(tid)]
+
+            pointers = process["pointers"]
+
+            if func not in NUM_ARGUMENTS:
+                raise Ignored("Unknown function")
+
+            if len(args) != NUM_ARGUMENTS[func]:
+                raise Ignored("Malformed input")
+
+            if func in ("jemalloc_stats", "free") and result:
+                raise Ignored("Malformed input")
+
+            if func in ("free", "realloc"):
+                ptr = int(args[0], 16)
+                if ptr and ptr not in pointers:
+                    raise Ignored("Did not see an alloc for pointer")
+                args[0] = "#%d" % pointers[ptr]
+                del pointers[ptr]
+
+            if result:
+                result = int(result, 16)
+                if not result:
+                    raise Ignored("Result is NULL")
+                result = "#%d" % pointers[result]
+
+            print(
+                "%d %d %s(%s)%s"
+                % (pid, tid, func, ",".join(args), "=%s" % result if result else "")
+            )
+
+        except Exception as e:
+            print('Ignored "%s": %s' % (line, e.message), file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/memory/replace/logalloc/replay/moz.build b/memory/replace/logalloc/replay/moz.build
new file mode 100644
index 0000000000..3e0c8c395a
--- /dev/null
+++ b/memory/replace/logalloc/replay/moz.build
@@ -0,0 +1,69 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+Program("logalloc-replay")
+
+SOURCES += [
+    "/mfbt/Assertions.cpp",
+    "/mfbt/Poison.cpp",
+    "/mfbt/RandomNum.cpp",
+    "/mfbt/TaggedAnonymousMemory.cpp",
+    "/mfbt/Unused.cpp",
+    "Replay.cpp",
+]
+
+if CONFIG["OS_TARGET"] == "Darwin":
+    # Work around "warning: 'aligned_alloc' is only available on macOS 10.15 or newer"
+    # when building with MACOSX_DEPLOYMENT_TARGET < 10.15 with >= 10.15 SDK.
+    # We have our own definition of the function, so it doesn't matter what the SDK says.
+    SOURCES["Replay.cpp"].flags += ["-Wno-unguarded-availability-new"]
+
+if CONFIG["MOZ_REPLACE_MALLOC_STATIC"] and (CONFIG["MOZ_DMD"] or CONFIG["MOZ_PHC"]):
+    UNIFIED_SOURCES += [
+        "/mfbt/HashFunctions.cpp",
+        "/mfbt/JSONWriter.cpp",
+        "/mozglue/misc/StackWalk.cpp",
+    ]
+    if CONFIG["OS_ARCH"] == "WINNT":
+        OS_LIBS += [
+            "dbghelp",
+        ]
+    if CONFIG["MOZ_LINKER"] and CONFIG["MOZ_WIDGET_TOOLKIT"] == "android":
+        LOCAL_INCLUDES += [
+            "/mozglue/linker",
+        ]
+        DEFINES["__wrap_dladdr"] = "dladdr"
+
+
+if CONFIG["MOZ_DMD"] or CONFIG["MOZ_PHC"]:
+    if CONFIG["MOZ_BUILD_APP"] == "memory":
+        EXPORTS.mozilla += [
+            "/mozglue/misc/StackWalk.h",
+        ]
+
+if not CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    SOURCES += [
+        "../FdPrintf.cpp",
+    ]
+
+LOCAL_INCLUDES += [
+    "..",
+]
+
+# Link replace-malloc and the default allocator.
+USE_LIBS += [
+    "memory",
+]
+
+# The memory library defines this, so it's needed here too.
+DEFINES["IMPL_MFBT"] = True
+
+if CONFIG["MOZ_NEEDS_LIBATOMIC"]:
+    OS_LIBS += ["atomic"]
+
+DisableStlWrapping()
+
+include("/mozglue/build/replace_malloc.mozbuild")
diff --git a/memory/replace/logalloc/replay/replay.log b/memory/replace/logalloc/replay/replay.log
new file mode 100644
index 0000000000..f1e6de788b
--- /dev/null
+++ b/memory/replace/logalloc/replay/replay.log
@@ -0,0 +1,18 @@
+1 1 jemalloc_stats()
+1 1 malloc(42)=#1
+1 1 malloc(24)=#2
+2 2 malloc(42)=#1
+1 1 free(#1)
+1 1 posix_memalign(4096,1024)=#1
+1 1 calloc(4,42)=#3
+1 1 free(#2)
+1 1 realloc(#3,84)=#2
+1 1 aligned_alloc(256,1024)=#3
+1 1 memalign(512,1024)=#4
+1 1 valloc(1024)=#5
+1 1 jemalloc_stats()
+1 1 free(#5)
+1 1 free(#4)
+1 1 free(#3)
+1 1 free(#2)
+1 1 free(#1)
diff --git a/memory/replace/moz.build b/memory/replace/moz.build
new file mode 100644
index 0000000000..a76c5aac58
--- /dev/null
+++ b/memory/replace/moz.build
@@ -0,0 +1,25 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+@template
+def ReplaceMalloc(name):
+    if CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+        DEFINES["MOZ_REPLACE_MALLOC_PREFIX"] = name.replace("-", "_")
+        FINAL_LIBRARY = "memory"
+    else:
+        SharedLibrary(name)
+
+
+DIRS += [
+    "logalloc",
+]
+
+if CONFIG["MOZ_DMD"]:
+    DIRS += ["dmd"]
+
+if CONFIG["MOZ_PHC"]:
+    DIRS += ["phc"]
diff --git a/memory/replace/phc/PHC.cpp b/memory/replace/phc/PHC.cpp
new file mode 100644
index 0000000000..7ddeac75b7
--- /dev/null
+++ b/memory/replace/phc/PHC.cpp
@@ -0,0 +1,1592 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// PHC is a probabilistic heap checker. A tiny fraction of randomly chosen heap
+// allocations are subject to some expensive checking via the use of OS page
+// access protection. A failed check triggers a crash, whereupon useful
+// information about the failure is put into the crash report. The cost and
+// coverage for each user is minimal, but spread over the entire user base the
+// coverage becomes significant.
+//
+// The idea comes from Chromium, where it is called GWP-ASAN. (Firefox uses PHC
+// as the name because GWP-ASAN is long, awkward, and doesn't have any
+// particular meaning.)
+//
+// In the current implementation up to 64 allocations per process can become
+// PHC allocations. These allocations must be page-sized or smaller. Each PHC
+// allocation gets its own page, and when the allocation is freed its page is
+// marked inaccessible until the page is reused for another allocation. This
+// means that a use-after-free defect (which includes double-frees) will be
+// caught if the use occurs before the page is reused for another allocation.
+// The crash report will contain stack traces for the allocation site, the free
+// site, and the use-after-free site, which is often enough to diagnose the
+// defect.
+//
+// Also, each PHC allocation is followed by a guard page. The PHC allocation is
+// positioned so that its end abuts the guard page (or as close as possible,
+// given alignment constraints). This means that a bounds violation at the end
+// of the allocation (overflow) will be caught. The crash report will contain
+// stack traces for the allocation site and the bounds violation use site,
+// which is often enough to diagnose the defect.
+//
+// (A bounds violation at the start of the allocation (underflow) will not be
+// caught, unless it is sufficiently large to hit the preceding allocation's
+// guard page, which is not that likely. It would be possible to look more
+// assiduously for underflow by randomly placing some allocations at the end of
+// the page and some at the start of the page, and GWP-ASAN does this. PHC does
+// not, however, because overflow is likely to be much more common than
+// underflow in practice.)
+//
+// We use a simple heuristic to categorize a guard page access as overflow or
+// underflow: if the address falls in the lower half of the guard page, we
+// assume it is overflow, otherwise we assume it is underflow. More
+// sophisticated heuristics are possible, but this one is very simple, and it is
+// likely that most overflows/underflows in practice are very close to the page
+// boundary.
+//
+// The design space for the randomization strategy is large. The current
+// implementation has a large random delay before it starts operating, and a
+// small random delay between each PHC allocation attempt. Each freed PHC
+// allocation is quarantined for a medium random delay before being reused, in
+// order to increase the chance of catching UAFs.
+//
+// The basic cost of PHC's operation is as follows.
+//
+// - The physical memory cost is 64 * 4 KiB = 256 KiB per process (assuming 4
+//   KiB pages) plus some metadata (including stack traces) for each page.
+//
+// - The virtual memory cost is the physical memory cost plus the guard pages:
+//   another 64 * 4 KiB = 256 KiB per process. PHC is currently only enabled on
+//   64-bit platforms so the impact of the virtual memory usage is negligible.
+//
+// - Every allocation requires a size check and a decrement-and-check of an
+//   atomic counter. When the counter reaches zero a PHC allocation can occur,
+//   which involves marking a page as accessible and getting a stack trace for
+//   the allocation site. Otherwise, mozjemalloc performs the allocation.
+//
+// - Every deallocation requires a range check on the pointer to see if it
+//   involves a PHC allocation. (The choice to only do PHC allocations that are
+//   a page or smaller enables this range check, because the 64 pages are
+//   contiguous. Allowing larger allocations would make this more complicated,
+//   and we definitely don't want something as slow as a hash table lookup on
+//   every deallocation.) PHC deallocations involve marking a page as
+//   inaccessible and getting a stack trace for the deallocation site.
+//
+// Note that calls to realloc(), free(), and malloc_usable_size() will
+// immediately crash if the given pointer falls within a page allocation's
+// page, but does not point to the start of the allocation itself.
+//
+//   void* p = malloc(64);
+//   free(p + 1);     // p+1 doesn't point to the allocation start; crash
+//
+// Such crashes will not have the PHC fields in the crash report.
+//
+// PHC-specific tests can be run with the following commands:
+// - gtests: `./mach gtest '*PHC*'`
+// - xpcshell-tests: `./mach test toolkit/crashreporter/test/unit`
+//   - This runs some non-PHC tests as well.
+
+#include "PHC.h"
+
+#include <stdlib.h>
+#include <time.h>
+
+#include <algorithm>
+
+#ifdef XP_WIN
+#  include <process.h>
+#else
+#  include <sys/mman.h>
+#  include <sys/types.h>
+#  include <pthread.h>
+#  include <unistd.h>
+#endif
+
+#include "replace_malloc.h"
+#include "FdPrintf.h"
+#include "Mutex.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/Atomics.h"
+#include "mozilla/Attributes.h"
+#include "mozilla/CheckedInt.h"
+#include "mozilla/Maybe.h"
+#include "mozilla/StackWalk.h"
+#include "mozilla/ThreadLocal.h"
+#include "mozilla/XorShift128PlusRNG.h"
+
+using namespace mozilla;
+
+//---------------------------------------------------------------------------
+// Utilities
+//---------------------------------------------------------------------------
+
+#ifdef ANDROID
+// Android doesn't have pthread_atfork defined in pthread.h.
+extern "C" MOZ_EXPORT int pthread_atfork(void (*)(void), void (*)(void),
+                                         void (*)(void));
+#endif
+
+#ifndef DISALLOW_COPY_AND_ASSIGN
+#  define DISALLOW_COPY_AND_ASSIGN(T) \
+    T(const T&);                      \
+    void operator=(const T&)
+#endif
+
+static malloc_table_t sMallocTable;
+
+// This class provides infallible operations for the small number of heap
+// allocations that PHC does for itself. It would be nice if we could use the
+// InfallibleAllocPolicy from mozalloc, but PHC cannot use mozalloc.
+class InfallibleAllocPolicy {
+ public:
+  static void AbortOnFailure(const void* aP) {
+    if (!aP) {
+      MOZ_CRASH("PHC failed to allocate");
+    }
+  }
+
+  template <class T>
+  static T* new_() {
+    void* p = sMallocTable.malloc(sizeof(T));
+    AbortOnFailure(p);
+    return new (p) T;
+  }
+};
+
+//---------------------------------------------------------------------------
+// Stack traces
+//---------------------------------------------------------------------------
+
+// This code is similar to the equivalent code within DMD.
+
+class StackTrace : public phc::StackTrace {
+ public:
+  StackTrace() : phc::StackTrace() {}
+
+  void Clear() { mLength = 0; }
+
+  void Fill();
+
+ private:
+  static void StackWalkCallback(uint32_t aFrameNumber, void* aPc, void* aSp,
+                                void* aClosure) {
+    StackTrace* st = (StackTrace*)aClosure;
+    MOZ_ASSERT(st->mLength < kMaxFrames);
+    st->mPcs[st->mLength] = aPc;
+    st->mLength++;
+    MOZ_ASSERT(st->mLength == aFrameNumber);
+  }
+};
+
+// WARNING WARNING WARNING: this function must only be called when GMut::sMutex
+// is *not* locked, otherwise we might get deadlocks.
+//
+// How? On Windows, MozStackWalk() can lock a mutex, M, from the shared library
+// loader. Another thread might call malloc() while holding M locked (when
+// loading a shared library) and try to lock GMut::sMutex, causing a deadlock.
+// So GMut::sMutex can't be locked during the call to MozStackWalk(). (For
+// details, see https://bugzilla.mozilla.org/show_bug.cgi?id=374829#c8. On
+// Linux, something similar can happen; see bug 824340. So we just disallow it
+// on all platforms.)
+//
+// In DMD, to avoid this problem we temporarily unlock the equivalent mutex for
+// the MozStackWalk() call. But that's grotty, and things are a bit different
+// here, so we just require that stack traces be obtained before locking
+// GMut::sMutex.
+//
+// Unfortunately, there is no reliable way at compile-time or run-time to ensure
+// this pre-condition. Hence this large comment.
+//
+void StackTrace::Fill() {
+  mLength = 0;
+
+#if defined(XP_WIN) && defined(_M_IX86)
+  // This avoids MozStackWalk(), which causes unusably slow startup on Win32
+  // when it is called during static initialization (see bug 1241684).
+  //
+  // This code is cribbed from the Gecko Profiler, which also uses
+  // FramePointerStackWalk() on Win32: Registers::SyncPopulate() for the
+  // frame pointer, and GetStackTop() for the stack end.
+  CONTEXT context;
+  RtlCaptureContext(&context);
+  void** fp = reinterpret_cast<void**>(context.Ebp);
+
+  PNT_TIB pTib = reinterpret_cast<PNT_TIB>(NtCurrentTeb());
+  void* stackEnd = static_cast<void*>(pTib->StackBase);
+  FramePointerStackWalk(StackWalkCallback, /* aSkipFrames = */ 0, kMaxFrames,
+                        this, fp, stackEnd);
+#elif defined(XP_MACOSX)
+  // This avoids MozStackWalk(), which has become unusably slow on Mac due to
+  // changes in libunwind.
+  //
+  // This code is cribbed from the Gecko Profiler, which also uses
+  // FramePointerStackWalk() on Mac: Registers::SyncPopulate() for the frame
+  // pointer, and GetStackTop() for the stack end.
+  void** fp;
+  asm(
+      // Dereference %rbp to get previous %rbp
+      "movq (%%rbp), %0\n\t"
+      : "=r"(fp));
+  void* stackEnd = pthread_get_stackaddr_np(pthread_self());
+  FramePointerStackWalk(StackWalkCallback, /* skipFrames = */ 0, kMaxFrames,
+                        this, fp, stackEnd);
+#else
+  MozStackWalk(StackWalkCallback, /* aSkipFrames = */ 0, kMaxFrames, this);
+#endif
+}
+
+//---------------------------------------------------------------------------
+// Logging
+//---------------------------------------------------------------------------
+
+// Change this to 1 to enable some PHC logging. Useful for debugging.
+#define PHC_LOGGING 0
+
+#if PHC_LOGGING
+
+static size_t GetPid() { return size_t(getpid()); }
+
+static size_t GetTid() {
+#  if defined(XP_WIN)
+  return size_t(GetCurrentThreadId());
+#  else
+  return size_t(pthread_self());
+#  endif
+}
+
+#  if defined(XP_WIN)
+#    define LOG_STDERR \
+      reinterpret_cast<intptr_t>(GetStdHandle(STD_ERROR_HANDLE))
+#  else
+#    define LOG_STDERR 2
+#  endif
+#  define LOG(fmt, ...)                                                \
+    FdPrintf(LOG_STDERR, "PHC[%zu,%zu,~%zu] " fmt, GetPid(), GetTid(), \
+             size_t(GAtomic::Now()), __VA_ARGS__)
+
+#else
+
+#  define LOG(fmt, ...)
+
+#endif  // PHC_LOGGING
+
+//---------------------------------------------------------------------------
+// Global state
+//---------------------------------------------------------------------------
+
+// Throughout this entire file time is measured as the number of sub-page
+// allocations performed (by PHC and mozjemalloc combined). `Time` is 64-bit
+// because we could have more than 2**32 allocations in a long-running session.
+// `Delay` is 32-bit because the delays used within PHC are always much smaller
+// than 2**32.
+using Time = uint64_t;   // A moment in time.
+using Delay = uint32_t;  // A time duration.
+
+// PHC only runs if the page size is 4 KiB; anything more is uncommon and would
+// use too much memory. So we hardwire this size.
+static const size_t kPageSize = 4096;
+
+// There are two kinds of page.
+// - Allocation pages, from which allocations are made.
+// - Guard pages, which are never touched by PHC.
+//
+// These page kinds are interleaved; each allocation page has a guard page on
+// either side.
+static const size_t kNumAllocPages = 64;
+static const size_t kNumAllPages = kNumAllocPages * 2 + 1;
+
+// The total size of the allocation pages and guard pages.
+static const size_t kAllPagesSize = kNumAllPages * kPageSize;
+
+// The junk value used to fill new allocation in debug builds. It's same value
+// as the one used by mozjemalloc. PHC applies it unconditionally in debug
+// builds. Unlike mozjemalloc, PHC doesn't consult the MALLOC_OPTIONS
+// environment variable to possibly change that behaviour.
+//
+// Also note that, unlike mozjemalloc, PHC doesn't have a poison value for freed
+// allocations because freed allocations are protected by OS page protection.
+const uint8_t kAllocJunk = 0xe4;
+
+// The maximum time.
+static const Time kMaxTime = ~(Time(0));
+
+// The average delay before doing any page allocations at the start of a
+// process. Note that roughly 1 million allocations occur in the main process
+// while starting the browser. The delay range is 1..kAvgFirstAllocDelay*2.
+static const Delay kAvgFirstAllocDelay = 512 * 1024;
+
+// The average delay until the next attempted page allocation, once we get past
+// the first delay. The delay range is 1..kAvgAllocDelay*2.
+static const Delay kAvgAllocDelay = 16 * 1024;
+
+// The average delay before reusing a freed page. Should be significantly larger
+// than kAvgAllocDelay, otherwise there's not much point in having it. The delay
+// range is (kAvgAllocDelay / 2)..(kAvgAllocDelay / 2 * 3). This is different to
+// the other delay ranges in not having a minimum of 1, because that's such a
+// short delay that there is a high likelihood of bad stacks in any crash
+// report.
+static const Delay kAvgPageReuseDelay = 256 * 1024;
+
+// Truncate aRnd to the range (1 .. AvgDelay*2). If aRnd is random, this
+// results in an average value of aAvgDelay + 0.5, which is close enough to
+// aAvgDelay. aAvgDelay must be a power-of-two (otherwise it will crash) for
+// speed.
+template <Delay AvgDelay>
+constexpr Delay Rnd64ToDelay(uint64_t aRnd) {
+  static_assert(IsPowerOfTwo(AvgDelay), "must be a power of two");
+
+  return aRnd % (AvgDelay * 2) + 1;
+}
+
+// Maps a pointer to a PHC-specific structure:
+// - Nothing
+// - A guard page (it is unspecified which one)
+// - An allocation page (with an index < kNumAllocPages)
+//
+// The standard way of handling a PtrKind is to check IsNothing(), and if that
+// fails, to check IsGuardPage(), and if that fails, to call AllocPage().
+class PtrKind {
+ private:
+  enum class Tag : uint8_t {
+    Nothing,
+    GuardPage,
+    AllocPage,
+  };
+
+  Tag mTag;
+  uintptr_t mIndex;  // Only used if mTag == Tag::AllocPage.
+
+ public:
+  // Detect what a pointer points to. This constructor must be fast because it
+  // is called for every call to free(), realloc(), malloc_usable_size(), and
+  // jemalloc_ptr_info().
+  PtrKind(const void* aPtr, const uint8_t* aPagesStart,
+          const uint8_t* aPagesLimit) {
+    if (!(aPagesStart <= aPtr && aPtr < aPagesLimit)) {
+      mTag = Tag::Nothing;
+    } else {
+      uintptr_t offset = static_cast<const uint8_t*>(aPtr) - aPagesStart;
+      uintptr_t allPageIndex = offset / kPageSize;
+      MOZ_ASSERT(allPageIndex < kNumAllPages);
+      if (allPageIndex & 1) {
+        // Odd-indexed pages are allocation pages.
+        uintptr_t allocPageIndex = allPageIndex / 2;
+        MOZ_ASSERT(allocPageIndex < kNumAllocPages);
+        mTag = Tag::AllocPage;
+        mIndex = allocPageIndex;
+      } else {
+        // Even-numbered pages are guard pages.
+        mTag = Tag::GuardPage;
+      }
+    }
+  }
+
+  bool IsNothing() const { return mTag == Tag::Nothing; }
+  bool IsGuardPage() const { return mTag == Tag::GuardPage; }
+
+  // This should only be called after IsNothing() and IsGuardPage() have been
+  // checked and failed.
+  uintptr_t AllocPageIndex() const {
+    MOZ_RELEASE_ASSERT(mTag == Tag::AllocPage);
+    return mIndex;
+  }
+};
+
+// Shared, atomic, mutable global state.
+class GAtomic {
+ public:
+  static void Init(Delay aFirstDelay) {
+    sAllocDelay = aFirstDelay;
+
+    LOG("Initial sAllocDelay <- %zu\n", size_t(aFirstDelay));
+  }
+
+  static Time Now() { return sNow; }
+
+  static void IncrementNow() { sNow++; }
+
+  // Decrements the delay and returns the decremented value.
+  static int32_t DecrementDelay() { return --sAllocDelay; }
+
+  static void SetAllocDelay(Delay aAllocDelay) { sAllocDelay = aAllocDelay; }
+
+ private:
+  // The current time. Relaxed semantics because it's primarily used for
+  // determining if an allocation can be recycled yet and therefore it doesn't
+  // need to be exact.
+  static Atomic<Time, Relaxed> sNow;
+
+  // Delay until the next attempt at a page allocation. See the comment in
+  // MaybePageAlloc() for an explanation of why it is a signed integer, and why
+  // it uses ReleaseAcquire semantics.
+  static Atomic<Delay, ReleaseAcquire> sAllocDelay;
+};
+
+Atomic<Time, Relaxed> GAtomic::sNow;
+Atomic<Delay, ReleaseAcquire> GAtomic::sAllocDelay;
+
+// Shared, immutable global state. Initialized by replace_init() and never
+// changed after that. replace_init() runs early enough that no synchronization
+// is needed.
+class GConst {
+ private:
+  // The bounds of the allocated pages.
+  uint8_t* const mPagesStart;
+  uint8_t* const mPagesLimit;
+
+  // Allocates the allocation pages and the guard pages, contiguously.
+  uint8_t* AllocAllPages() {
+    // Allocate the pages so that they are inaccessible. They are never freed,
+    // because it would happen at process termination when it would be of little
+    // use.
+    void* pages =
+#ifdef XP_WIN
+        VirtualAlloc(nullptr, kAllPagesSize, MEM_RESERVE, PAGE_NOACCESS);
+#else
+        mmap(nullptr, kAllPagesSize, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE, -1,
+             0);
+#endif
+    if (!pages) {
+      MOZ_CRASH();
+    }
+
+    return static_cast<uint8_t*>(pages);
+  }
+
+ public:
+  GConst()
+      : mPagesStart(AllocAllPages()), mPagesLimit(mPagesStart + kAllPagesSize) {
+    LOG("AllocAllPages at %p..%p\n", mPagesStart, mPagesLimit);
+  }
+
+  class PtrKind PtrKind(const void* aPtr) {
+    class PtrKind pk(aPtr, mPagesStart, mPagesLimit);
+    return pk;
+  }
+
+  bool IsInFirstGuardPage(const void* aPtr) {
+    return mPagesStart <= aPtr && aPtr < mPagesStart + kPageSize;
+  }
+
+  // Get the address of the allocation page referred to via an index. Used when
+  // marking the page as accessible/inaccessible.
+  uint8_t* AllocPagePtr(uintptr_t aIndex) {
+    MOZ_ASSERT(aIndex < kNumAllocPages);
+    // Multiply by two and add one to account for allocation pages *and* guard
+    // pages.
+    return mPagesStart + (2 * aIndex + 1) * kPageSize;
+  }
+};
+
+static GConst* gConst;
+
+// On MacOS, the first __thread/thread_local access calls malloc, which leads
+// to an infinite loop. So we use pthread-based TLS instead, which somehow
+// doesn't have this problem.
+#if !defined(XP_DARWIN)
+#  define PHC_THREAD_LOCAL(T) MOZ_THREAD_LOCAL(T)
+#else
+#  define PHC_THREAD_LOCAL(T) \
+    detail::ThreadLocal<T, detail::ThreadLocalKeyStorage>
+#endif
+
+// Thread-local state.
+class GTls {
+  GTls(const GTls&) = delete;
+
+  const GTls& operator=(const GTls&) = delete;
+
+  // When true, PHC does as little as possible.
+  //
+  // (a) It does not allocate any new page allocations.
+  //
+  // (b) It avoids doing any operations that might call malloc/free/etc., which
+  //     would cause re-entry into PHC. (In practice, MozStackWalk() is the
+  //     only such operation.) Note that calls to the functions in sMallocTable
+  //     are ok.
+  //
+  // For example, replace_malloc() will just fall back to mozjemalloc. However,
+  // operations involving existing allocations are more complex, because those
+  // existing allocations may be page allocations. For example, if
+  // replace_free() is passed a page allocation on a PHC-disabled thread, it
+  // will free the page allocation in the usual way, but it will get a dummy
+  // freeStack in order to avoid calling MozStackWalk(), as per (b) above.
+  //
+  // This single disabling mechanism has two distinct uses.
+  //
+  // - It's used to prevent re-entry into PHC, which can cause correctness
+  //   problems. For example, consider this sequence.
+  //
+  //   1. enter replace_free()
+  //   2. which calls PageFree()
+  //   3. which calls MozStackWalk()
+  //   4. which locks a mutex M, and then calls malloc
+  //   5. enter replace_malloc()
+  //   6. which calls MaybePageAlloc()
+  //   7. which calls MozStackWalk()
+  //   8. which (re)locks a mutex M --> deadlock
+  //
+  //   We avoid this sequence by "disabling" the thread in PageFree() (at step
+  //   2), which causes MaybePageAlloc() to fail, avoiding the call to
+  //   MozStackWalk() (at step 7).
+  //
+  //   In practice, realloc or free of a PHC allocation is unlikely on a thread
+  //   that is disabled because of this use: MozStackWalk() will probably only
+  //   realloc/free allocations that it allocated itself, but those won't be
+  //   page allocations because PHC is disabled before calling MozStackWalk().
+  //
+  //   (Note that MaybePageAlloc() could safely do a page allocation so long as
+  //   it avoided calling MozStackWalk() by getting a dummy allocStack. But it
+  //   wouldn't be useful, and it would prevent the second use below.)
+  //
+  // - It's used to prevent PHC allocations in some tests that rely on
+  //   mozjemalloc's exact allocation behaviour, which PHC does not replicate
+  //   exactly. (Note that (b) isn't necessary for this use -- MozStackWalk()
+  //   could be safely called -- but it is necessary for the first use above.)
+  //
+  static PHC_THREAD_LOCAL(bool) tlsIsDisabled;
+
+ public:
+  static void Init() {
+    if (!tlsIsDisabled.init()) {
+      MOZ_CRASH();
+    }
+  }
+
+  static void DisableOnCurrentThread() {
+    MOZ_ASSERT(!GTls::tlsIsDisabled.get());
+    tlsIsDisabled.set(true);
+  }
+
+  static void EnableOnCurrentThread() {
+    MOZ_ASSERT(GTls::tlsIsDisabled.get());
+    tlsIsDisabled.set(false);
+  }
+
+  static bool IsDisabledOnCurrentThread() { return tlsIsDisabled.get(); }
+};
+
+PHC_THREAD_LOCAL(bool) GTls::tlsIsDisabled;
+
+class AutoDisableOnCurrentThread {
+  AutoDisableOnCurrentThread(const AutoDisableOnCurrentThread&) = delete;
+
+  const AutoDisableOnCurrentThread& operator=(
+      const AutoDisableOnCurrentThread&) = delete;
+
+ public:
+  explicit AutoDisableOnCurrentThread() { GTls::DisableOnCurrentThread(); }
+  ~AutoDisableOnCurrentThread() { GTls::EnableOnCurrentThread(); }
+};
+
+// This type is used as a proof-of-lock token, to make it clear which functions
+// require sMutex to be locked.
+using GMutLock = const MutexAutoLock&;
+
+// Shared, mutable global state. Protected by sMutex; all accessing functions
+// take a GMutLock as proof that sMutex is held.
+class GMut {
+  enum class AllocPageState {
+    NeverAllocated = 0,
+    InUse = 1,
+    Freed = 2,
+  };
+
+  // Metadata for each allocation page.
+  class AllocPageInfo {
+   public:
+    AllocPageInfo()
+        : mState(AllocPageState::NeverAllocated),
+          mArenaId(),
+          mBaseAddr(nullptr),
+          mAllocStack(),
+          mFreeStack(),
+          mReuseTime(0) {}
+
+    // The current allocation page state.
+    AllocPageState mState;
+
+    // The arena that the allocation is nominally from. This isn't meaningful
+    // within PHC, which has no arenas. But it is necessary for reallocation of
+    // page allocations as normal allocations, such as in this code:
+    //
+    //   p = moz_arena_malloc(arenaId, 4096);
+    //   realloc(p, 8192);
+    //
+    // The realloc is more than one page, and thus too large for PHC to handle.
+    // Therefore, if PHC handles the first allocation, it must ask mozjemalloc
+    // to allocate the 8192 bytes in the correct arena, and to do that, it must
+    // call sMallocTable.moz_arena_malloc with the correct arenaId under the
+    // covers. Therefore it must record that arenaId.
+    //
+    // This field is also needed for jemalloc_ptr_info() to work, because it
+    // also returns the arena ID (but only in debug builds).
+    //
+    // - NeverAllocated: must be 0.
+    // - InUse | Freed: can be any valid arena ID value.
+    Maybe<arena_id_t> mArenaId;
+
+    // The starting address of the allocation. Will not be the same as the page
+    // address unless the allocation is a full page.
+    // - NeverAllocated: must be 0.
+    // - InUse | Freed: must be within the allocation page.
+    uint8_t* mBaseAddr;
+
+    // Usable size is computed as the number of bytes between the pointer and
+    // the end of the allocation page. This might be bigger than the requested
+    // size, especially if an outsized alignment is requested.
+    size_t UsableSize() const {
+      return mState == AllocPageState::NeverAllocated
+                 ? 0
+                 : kPageSize - (reinterpret_cast<uintptr_t>(mBaseAddr) &
+                                (kPageSize - 1));
+    }
+
+    // The allocation stack.
+    // - NeverAllocated: Nothing.
+    // - InUse | Freed: Some.
+    Maybe<StackTrace> mAllocStack;
+
+    // The free stack.
+    // - NeverAllocated | InUse: Nothing.
+    // - Freed: Some.
+    Maybe<StackTrace> mFreeStack;
+
+    // The time at which the page is available for reuse, as measured against
+    // GAtomic::sNow. When the page is in use this value will be kMaxTime.
+    // - NeverAllocated: must be 0.
+    // - InUse: must be kMaxTime.
+    // - Freed: must be > 0 and < kMaxTime.
+    Time mReuseTime;
+  };
+
+ public:
+  // The mutex that protects the other members.
+  static Mutex sMutex;
+
+  GMut()
+      : mRNG(RandomSeed<0>(), RandomSeed<1>()),
+        mAllocPages(),
+        mNumPageAllocs(0),
+        mPageAllocHits(0),
+        mPageAllocMisses(0) {
+    sMutex.Init();
+  }
+
+  uint64_t Random64(GMutLock) { return mRNG.next(); }
+
+  bool IsPageInUse(GMutLock, uintptr_t aIndex) {
+    return mAllocPages[aIndex].mState == AllocPageState::InUse;
+  }
+
+  // Is the page free? And if so, has enough time passed that we can use it?
+  bool IsPageAllocatable(GMutLock, uintptr_t aIndex, Time aNow) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    return page.mState != AllocPageState::InUse && aNow >= page.mReuseTime;
+  }
+
+  Maybe<arena_id_t> PageArena(GMutLock aLock, uintptr_t aIndex) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    return page.mArenaId;
+  }
+
+  size_t PageUsableSize(GMutLock aLock, uintptr_t aIndex) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    return page.UsableSize();
+  }
+
+  void SetPageInUse(GMutLock aLock, uintptr_t aIndex,
+                    const Maybe<arena_id_t>& aArenaId, uint8_t* aBaseAddr,
+                    const StackTrace& aAllocStack) {
+    AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageNotInUse(aLock, page);
+
+    page.mState = AllocPageState::InUse;
+    page.mArenaId = aArenaId;
+    page.mBaseAddr = aBaseAddr;
+    page.mAllocStack = Some(aAllocStack);
+    page.mFreeStack = Nothing();
+    page.mReuseTime = kMaxTime;
+
+    mNumPageAllocs++;
+    MOZ_RELEASE_ASSERT(mNumPageAllocs <= kNumAllocPages);
+  }
+
+  void ResizePageInUse(GMutLock aLock, uintptr_t aIndex,
+                       const Maybe<arena_id_t>& aArenaId, uint8_t* aNewBaseAddr,
+                       const StackTrace& aAllocStack) {
+    AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    // page.mState is not changed.
+    if (aArenaId.isSome()) {
+      // Crash if the arenas don't match.
+      MOZ_RELEASE_ASSERT(page.mArenaId == aArenaId);
+    }
+    page.mBaseAddr = aNewBaseAddr;
+    // We could just keep the original alloc stack, but the realloc stack is
+    // more recent and therefore seems more useful.
+    page.mAllocStack = Some(aAllocStack);
+    // page.mFreeStack is not changed.
+    // page.mReuseTime is not changed.
+  };
+
+  void SetPageFreed(GMutLock aLock, uintptr_t aIndex,
+                    const Maybe<arena_id_t>& aArenaId,
+                    const StackTrace& aFreeStack, Delay aReuseDelay) {
+    AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    page.mState = AllocPageState::Freed;
+
+    // page.mArenaId is left unchanged, for jemalloc_ptr_info() calls that
+    // occur after freeing (e.g. in the PtrInfo test in TestJemalloc.cpp).
+    if (aArenaId.isSome()) {
+      // Crash if the arenas don't match.
+      MOZ_RELEASE_ASSERT(page.mArenaId == aArenaId);
+    }
+
+    // page.musableSize is left unchanged, for reporting on UAF, and for
+    // jemalloc_ptr_info() calls that occur after freeing (e.g. in the PtrInfo
+    // test in TestJemalloc.cpp).
+
+    // page.mAllocStack is left unchanged, for reporting on UAF.
+
+    page.mFreeStack = Some(aFreeStack);
+    page.mReuseTime = GAtomic::Now() + aReuseDelay;
+
+    MOZ_RELEASE_ASSERT(mNumPageAllocs > 0);
+    mNumPageAllocs--;
+  }
+
+  static void CrashOnGuardPage(void* aPtr) {
+    // An operation on a guard page? This is a bounds violation. Deliberately
+    // touch the page in question, to cause a crash that triggers the usual PHC
+    // machinery.
+    LOG("CrashOnGuardPage(%p), bounds violation\n", aPtr);
+    *static_cast<uint8_t*>(aPtr) = 0;
+    MOZ_CRASH("unreachable");
+  }
+
+  void EnsureValidAndInUse(GMutLock, void* aPtr, uintptr_t aIndex) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+
+    // The pointer must point to the start of the allocation.
+    MOZ_RELEASE_ASSERT(page.mBaseAddr == aPtr);
+
+    if (page.mState == AllocPageState::Freed) {
+      // An operation on a freed page? This is a particular kind of
+      // use-after-free. Deliberately touch the page in question, in order to
+      // cause a crash that triggers the usual PHC machinery. But unlock sMutex
+      // first, because that self-same PHC machinery needs to re-lock it, and
+      // the crash causes non-local control flow so sMutex won't be unlocked
+      // the normal way in the caller.
+      LOG("EnsureValidAndInUse(%p), use-after-free\n", aPtr);
+      sMutex.Unlock();
+      *static_cast<uint8_t*>(aPtr) = 0;
+      MOZ_CRASH("unreachable");
+    }
+  }
+
+  void FillAddrInfo(GMutLock, uintptr_t aIndex, const void* aBaseAddr,
+                    bool isGuardPage, phc::AddrInfo& aOut) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    if (isGuardPage) {
+      aOut.mKind = phc::AddrInfo::Kind::GuardPage;
+    } else {
+      switch (page.mState) {
+        case AllocPageState::NeverAllocated:
+          aOut.mKind = phc::AddrInfo::Kind::NeverAllocatedPage;
+          break;
+
+        case AllocPageState::InUse:
+          aOut.mKind = phc::AddrInfo::Kind::InUsePage;
+          break;
+
+        case AllocPageState::Freed:
+          aOut.mKind = phc::AddrInfo::Kind::FreedPage;
+          break;
+
+        default:
+          MOZ_CRASH();
+      }
+    }
+    aOut.mBaseAddr = page.mBaseAddr;
+    aOut.mUsableSize = page.UsableSize();
+    aOut.mAllocStack = page.mAllocStack;
+    aOut.mFreeStack = page.mFreeStack;
+  }
+
+  void FillJemallocPtrInfo(GMutLock, const void* aPtr, uintptr_t aIndex,
+                           jemalloc_ptr_info_t* aInfo) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    switch (page.mState) {
+      case AllocPageState::NeverAllocated:
+        break;
+
+      case AllocPageState::InUse: {
+        // Only return TagLiveAlloc if the pointer is within the bounds of the
+        // allocation's usable size.
+        uint8_t* base = page.mBaseAddr;
+        uint8_t* limit = base + page.UsableSize();
+        if (base <= aPtr && aPtr < limit) {
+          *aInfo = {TagLiveAlloc, page.mBaseAddr, page.UsableSize(),
+                    page.mArenaId.valueOr(0)};
+          return;
+        }
+        break;
+      }
+
+      case AllocPageState::Freed: {
+        // Only return TagFreedAlloc if the pointer is within the bounds of the
+        // former allocation's usable size.
+        uint8_t* base = page.mBaseAddr;
+        uint8_t* limit = base + page.UsableSize();
+        if (base <= aPtr && aPtr < limit) {
+          *aInfo = {TagFreedAlloc, page.mBaseAddr, page.UsableSize(),
+                    page.mArenaId.valueOr(0)};
+          return;
+        }
+        break;
+      }
+
+      default:
+        MOZ_CRASH();
+    }
+
+    // Pointers into guard pages will end up here, as will pointers into
+    // allocation pages that aren't within the allocation's bounds.
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+  }
+
+  static void prefork() { sMutex.Lock(); }
+  static void postfork() { sMutex.Unlock(); }
+
+  void IncPageAllocHits(GMutLock) { mPageAllocHits++; }
+  void IncPageAllocMisses(GMutLock) { mPageAllocMisses++; }
+
+  size_t NumPageAllocs(GMutLock) { return mNumPageAllocs; }
+
+  size_t PageAllocHits(GMutLock) { return mPageAllocHits; }
+  size_t PageAllocAttempts(GMutLock) {
+    return mPageAllocHits + mPageAllocMisses;
+  }
+
+  // This is an integer because FdPrintf only supports integer printing.
+  size_t PageAllocHitRate(GMutLock) {
+    return mPageAllocHits * 100 / (mPageAllocHits + mPageAllocMisses);
+  }
+
+ private:
+  template <int N>
+  uint64_t RandomSeed() {
+    // An older version of this code used RandomUint64() here, but on Mac that
+    // function uses arc4random(), which can allocate, which would cause
+    // re-entry, which would be bad. So we just use time() and a local variable
+    // address. These are mediocre sources of entropy, but good enough for PHC.
+    static_assert(N == 0 || N == 1, "must be 0 or 1");
+    uint64_t seed;
+    if (N == 0) {
+      time_t t = time(nullptr);
+      seed = t ^ (t << 32);
+    } else {
+      seed = uintptr_t(&seed) ^ (uintptr_t(&seed) << 32);
+    }
+    return seed;
+  }
+
+  void AssertAllocPageInUse(GMutLock, const AllocPageInfo& aPage) {
+    MOZ_ASSERT(aPage.mState == AllocPageState::InUse);
+    // There is nothing to assert about aPage.mArenaId.
+    MOZ_ASSERT(aPage.mBaseAddr);
+    MOZ_ASSERT(aPage.UsableSize() > 0);
+    MOZ_ASSERT(aPage.mAllocStack.isSome());
+    MOZ_ASSERT(aPage.mFreeStack.isNothing());
+    MOZ_ASSERT(aPage.mReuseTime == kMaxTime);
+  }
+
+  void AssertAllocPageNotInUse(GMutLock, const AllocPageInfo& aPage) {
+    // We can assert a lot about `NeverAllocated` pages, but not much about
+    // `Freed` pages.
+#ifdef DEBUG
+    bool isFresh = aPage.mState == AllocPageState::NeverAllocated;
+    MOZ_ASSERT(isFresh || aPage.mState == AllocPageState::Freed);
+    MOZ_ASSERT_IF(isFresh, aPage.mArenaId == Nothing());
+    MOZ_ASSERT(isFresh == (aPage.mBaseAddr == nullptr));
+    MOZ_ASSERT(isFresh == (aPage.mAllocStack.isNothing()));
+    MOZ_ASSERT(isFresh == (aPage.mFreeStack.isNothing()));
+    MOZ_ASSERT(aPage.mReuseTime != kMaxTime);
+#endif
+  }
+
+  // RNG for deciding which allocations to treat specially. It doesn't need to
+  // be high quality.
+  //
+  // This is a raw pointer for the reason explained in the comment above
+  // GMut's constructor. Don't change it to UniquePtr or anything like that.
+  non_crypto::XorShift128PlusRNG mRNG;
+
+  AllocPageInfo mAllocPages[kNumAllocPages];
+
+  // How many page allocs are currently in use (the max is kNumAllocPages).
+  size_t mNumPageAllocs;
+
+  // How many allocations that could have been page allocs actually were? As
+  // constrained kNumAllocPages. If the hit ratio isn't close to 100% it's
+  // likely that the global constants are poorly chosen.
+  size_t mPageAllocHits;
+  size_t mPageAllocMisses;
+};
+
+Mutex GMut::sMutex;
+
+static GMut* gMut;
+
+//---------------------------------------------------------------------------
+// Page allocation operations
+//---------------------------------------------------------------------------
+
+// Attempt a page allocation if the time and the size are right. Allocated
+// memory is zeroed if aZero is true. On failure, the caller should attempt a
+// normal allocation via sMallocTable. Can be called in a context where
+// GMut::sMutex is locked.
+static void* MaybePageAlloc(const Maybe<arena_id_t>& aArenaId, size_t aReqSize,
+                            size_t aAlignment, bool aZero) {
+  MOZ_ASSERT(IsPowerOfTwo(aAlignment));
+
+  if (aReqSize > kPageSize) {
+    return nullptr;
+  }
+
+  GAtomic::IncrementNow();
+
+  // Decrement the delay. If it's zero, we do a page allocation and reset the
+  // delay to a random number. Because the assignment to the random number isn't
+  // atomic w.r.t. the decrement, we might have a sequence like this:
+  //
+  //     Thread 1                      Thread 2           Thread 3
+  //     --------                      --------           --------
+  // (a) newDelay = --sAllocDelay (-> 0)
+  // (b)                               --sAllocDelay (-> -1)
+  // (c) (newDelay != 0) fails
+  // (d)                                                  --sAllocDelay (-> -2)
+  // (e) sAllocDelay = new_random_number()
+  //
+  // It's critical that sAllocDelay has ReleaseAcquire semantics, because that
+  // guarantees that exactly one thread will see sAllocDelay have the value 0.
+  // (Relaxed semantics wouldn't guarantee that.)
+  //
+  // It's also nice that sAllocDelay is signed, given that we can decrement to
+  // below zero. (Strictly speaking, an unsigned integer would also work due
+  // to wrapping, but a signed integer is conceptually cleaner.)
+  //
+  // Finally, note that the decrements that occur between (a) and (e) above are
+  // effectively ignored, because (e) clobbers them. This shouldn't be a
+  // problem; it effectively just adds a little more randomness to
+  // new_random_number(). An early version of this code tried to account for
+  // these decrements by doing `sAllocDelay += new_random_number()`. However, if
+  // new_random_value() is small, the number of decrements between (a) and (e)
+  // can easily exceed it, whereupon sAllocDelay ends up negative after
+  // `sAllocDelay += new_random_number()`, and the zero-check never succeeds
+  // again. (At least, not until sAllocDelay wraps around on overflow, which
+  // would take a very long time indeed.)
+  //
+  int32_t newDelay = GAtomic::DecrementDelay();
+  if (newDelay != 0) {
+    return nullptr;
+  }
+
+  if (GTls::IsDisabledOnCurrentThread()) {
+    return nullptr;
+  }
+
+  // Disable on this thread *before* getting the stack trace.
+  AutoDisableOnCurrentThread disable;
+
+  // Get the stack trace *before* locking the mutex. If we return nullptr then
+  // it was a waste, but it's not so frequent, and doing a stack walk while
+  // the mutex is locked is problematic (see the big comment on
+  // StackTrace::Fill() for details).
+  StackTrace allocStack;
+  allocStack.Fill();
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  Time now = GAtomic::Now();
+  Delay newAllocDelay = Rnd64ToDelay<kAvgAllocDelay>(gMut->Random64(lock));
+
+  // We start at a random page alloc and wrap around, to ensure pages get even
+  // amounts of use.
+  uint8_t* ptr = nullptr;
+  uint8_t* pagePtr = nullptr;
+  for (uintptr_t n = 0, i = size_t(gMut->Random64(lock)) % kNumAllocPages;
+       n < kNumAllocPages; n++, i = (i + 1) % kNumAllocPages) {
+    if (!gMut->IsPageAllocatable(lock, i, now)) {
+      continue;
+    }
+
+    pagePtr = gConst->AllocPagePtr(i);
+    MOZ_ASSERT(pagePtr);
+    bool ok =
+#ifdef XP_WIN
+        !!VirtualAlloc(pagePtr, kPageSize, MEM_COMMIT, PAGE_READWRITE);
+#else
+        mprotect(pagePtr, kPageSize, PROT_READ | PROT_WRITE) == 0;
+#endif
+    size_t usableSize = sMallocTable.malloc_good_size(aReqSize);
+    if (ok) {
+      MOZ_ASSERT(usableSize > 0);
+
+      // Put the allocation as close to the end of the page as possible,
+      // allowing for alignment requirements.
+      ptr = pagePtr + kPageSize - usableSize;
+      if (aAlignment != 1) {
+        ptr = reinterpret_cast<uint8_t*>(
+            (reinterpret_cast<uintptr_t>(ptr) & ~(aAlignment - 1)));
+      }
+
+      gMut->SetPageInUse(lock, i, aArenaId, ptr, allocStack);
+
+      if (aZero) {
+        memset(ptr, 0, usableSize);
+      } else {
+#ifdef DEBUG
+        memset(ptr, kAllocJunk, usableSize);
+#endif
+      }
+    }
+
+    gMut->IncPageAllocHits(lock);
+    LOG("PageAlloc(%zu, %zu) -> %p[%zu]/%p (%zu) (z%zu), sAllocDelay <- %zu, "
+        "fullness %zu/%zu, hits %zu/%zu (%zu%%)\n",
+        aReqSize, aAlignment, pagePtr, i, ptr, usableSize, size_t(aZero),
+        size_t(newAllocDelay), gMut->NumPageAllocs(lock), kNumAllocPages,
+        gMut->PageAllocHits(lock), gMut->PageAllocAttempts(lock),
+        gMut->PageAllocHitRate(lock));
+    break;
+  }
+
+  if (!pagePtr) {
+    // No pages are available, or VirtualAlloc/mprotect failed.
+    gMut->IncPageAllocMisses(lock);
+    LOG("No PageAlloc(%zu, %zu), sAllocDelay <- %zu, fullness %zu/%zu, hits "
+        "%zu/%zu "
+        "(%zu%%)\n",
+        aReqSize, aAlignment, size_t(newAllocDelay), gMut->NumPageAllocs(lock),
+        kNumAllocPages, gMut->PageAllocHits(lock),
+        gMut->PageAllocAttempts(lock), gMut->PageAllocHitRate(lock));
+  }
+
+  // Set the new alloc delay.
+  GAtomic::SetAllocDelay(newAllocDelay);
+
+  return ptr;
+}
+
+static void FreePage(GMutLock aLock, uintptr_t aIndex,
+                     const Maybe<arena_id_t>& aArenaId,
+                     const StackTrace& aFreeStack, Delay aReuseDelay) {
+  void* pagePtr = gConst->AllocPagePtr(aIndex);
+#ifdef XP_WIN
+  if (!VirtualFree(pagePtr, kPageSize, MEM_DECOMMIT)) {
+    return;
+  }
+#else
+  if (!mmap(pagePtr, kPageSize, PROT_NONE, MAP_FIXED | MAP_PRIVATE | MAP_ANON,
+            -1, 0)) {
+    return;
+  }
+#endif
+
+  gMut->SetPageFreed(aLock, aIndex, aArenaId, aFreeStack, aReuseDelay);
+}
+
+//---------------------------------------------------------------------------
+// replace-malloc machinery
+//---------------------------------------------------------------------------
+
+// This handles malloc, moz_arena_malloc, and realloc-with-a-nullptr.
+MOZ_ALWAYS_INLINE static void* PageMalloc(const Maybe<arena_id_t>& aArenaId,
+                                          size_t aReqSize) {
+  void* ptr = MaybePageAlloc(aArenaId, aReqSize, /* aAlignment */ 1,
+                             /* aZero */ false);
+  return ptr ? ptr
+             : (aArenaId.isSome()
+                    ? sMallocTable.moz_arena_malloc(*aArenaId, aReqSize)
+                    : sMallocTable.malloc(aReqSize));
+}
+
+static void* replace_malloc(size_t aReqSize) {
+  return PageMalloc(Nothing(), aReqSize);
+}
+
+static Delay ReuseDelay(GMutLock aLock) {
+  return (kAvgPageReuseDelay / 2) +
+         Rnd64ToDelay<kAvgPageReuseDelay / 2>(gMut->Random64(aLock));
+}
+
+// This handles both calloc and moz_arena_calloc.
+MOZ_ALWAYS_INLINE static void* PageCalloc(const Maybe<arena_id_t>& aArenaId,
+                                          size_t aNum, size_t aReqSize) {
+  CheckedInt<size_t> checkedSize = CheckedInt<size_t>(aNum) * aReqSize;
+  if (!checkedSize.isValid()) {
+    return nullptr;
+  }
+
+  void* ptr = MaybePageAlloc(aArenaId, checkedSize.value(), /* aAlignment */ 1,
+                             /* aZero */ true);
+  return ptr ? ptr
+             : (aArenaId.isSome()
+                    ? sMallocTable.moz_arena_calloc(*aArenaId, aNum, aReqSize)
+                    : sMallocTable.calloc(aNum, aReqSize));
+}
+
+static void* replace_calloc(size_t aNum, size_t aReqSize) {
+  return PageCalloc(Nothing(), aNum, aReqSize);
+}
+
+// This function handles both realloc and moz_arena_realloc.
+//
+// As always, realloc is complicated, and doubly so when there are two
+// different kinds of allocations in play. Here are the possible transitions,
+// and what we do in practice.
+//
+// - normal-to-normal: This is straightforward and obviously necessary.
+//
+// - normal-to-page: This is disallowed because it would require getting the
+//   arenaId of the normal allocation, which isn't possible in non-DEBUG builds
+//   for security reasons.
+//
+// - page-to-page: This is done whenever possible, i.e. whenever the new size
+//   is less than or equal to 4 KiB. This choice counterbalances the
+//   disallowing of normal-to-page allocations, in order to avoid biasing
+//   towards or away from page allocations. It always occurs in-place.
+//
+// - page-to-normal: this is done only when necessary, i.e. only when the new
+//   size is greater than 4 KiB. This choice naturally flows from the
+//   prior choice on page-to-page transitions.
+//
+// In summary: realloc doesn't change the allocation kind unless it must.
+//
+MOZ_ALWAYS_INLINE static void* PageRealloc(const Maybe<arena_id_t>& aArenaId,
+                                           void* aOldPtr, size_t aNewSize) {
+  if (!aOldPtr) {
+    // Null pointer. Treat like malloc(aNewSize).
+    return PageMalloc(aArenaId, aNewSize);
+  }
+
+  PtrKind pk = gConst->PtrKind(aOldPtr);
+  if (pk.IsNothing()) {
+    // A normal-to-normal transition.
+    return aArenaId.isSome()
+               ? sMallocTable.moz_arena_realloc(*aArenaId, aOldPtr, aNewSize)
+               : sMallocTable.realloc(aOldPtr, aNewSize);
+  }
+
+  if (pk.IsGuardPage()) {
+    GMut::CrashOnGuardPage(aOldPtr);
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  // A page-to-something transition.
+
+  // Note that `disable` has no effect unless it is emplaced below.
+  Maybe<AutoDisableOnCurrentThread> disable;
+  // Get the stack trace *before* locking the mutex.
+  StackTrace stack;
+  if (GTls::IsDisabledOnCurrentThread()) {
+    // PHC is disabled on this thread. Leave the stack empty.
+  } else {
+    // Disable on this thread *before* getting the stack trace.
+    disable.emplace();
+    stack.Fill();
+  }
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  // Check for realloc() of a freed block.
+  gMut->EnsureValidAndInUse(lock, aOldPtr, index);
+
+  if (aNewSize <= kPageSize) {
+    // A page-to-page transition. Just keep using the page allocation. We do
+    // this even if the thread is disabled, because it doesn't create a new
+    // page allocation. Note that ResizePageInUse() checks aArenaId.
+    //
+    // Move the bytes with memmove(), because the old allocation and the new
+    // allocation overlap. Move the usable size rather than the requested size,
+    // because the user might have used malloc_usable_size() and filled up the
+    // usable size.
+    size_t oldUsableSize = gMut->PageUsableSize(lock, index);
+    size_t newUsableSize = sMallocTable.malloc_good_size(aNewSize);
+    uint8_t* pagePtr = gConst->AllocPagePtr(index);
+    uint8_t* newPtr = pagePtr + kPageSize - newUsableSize;
+    memmove(newPtr, aOldPtr, std::min(oldUsableSize, aNewSize));
+    gMut->ResizePageInUse(lock, index, aArenaId, newPtr, stack);
+    LOG("PageRealloc-Reuse(%p, %zu) -> %p\n", aOldPtr, aNewSize, newPtr);
+    return newPtr;
+  }
+
+  // A page-to-normal transition (with the new size greater than page-sized).
+  // (Note that aArenaId is checked below.)
+  void* newPtr;
+  if (aArenaId.isSome()) {
+    newPtr = sMallocTable.moz_arena_malloc(*aArenaId, aNewSize);
+  } else {
+    Maybe<arena_id_t> oldArenaId = gMut->PageArena(lock, index);
+    newPtr = (oldArenaId.isSome()
+                  ? sMallocTable.moz_arena_malloc(*oldArenaId, aNewSize)
+                  : sMallocTable.malloc(aNewSize));
+  }
+  if (!newPtr) {
+    return nullptr;
+  }
+
+  MOZ_ASSERT(aNewSize > kPageSize);
+
+  Delay reuseDelay = ReuseDelay(lock);
+
+  // Copy the usable size rather than the requested size, because the user
+  // might have used malloc_usable_size() and filled up the usable size. Note
+  // that FreePage() checks aArenaId (via SetPageFreed()).
+  size_t oldUsableSize = gMut->PageUsableSize(lock, index);
+  memcpy(newPtr, aOldPtr, std::min(oldUsableSize, aNewSize));
+  FreePage(lock, index, aArenaId, stack, reuseDelay);
+  LOG("PageRealloc-Free(%p[%zu], %zu) -> %p, %zu delay, reuse at ~%zu\n",
+      aOldPtr, index, aNewSize, newPtr, size_t(reuseDelay),
+      size_t(GAtomic::Now()) + reuseDelay);
+
+  return newPtr;
+}
+
+static void* replace_realloc(void* aOldPtr, size_t aNewSize) {
+  return PageRealloc(Nothing(), aOldPtr, aNewSize);
+}
+
+// This handles both free and moz_arena_free.
+MOZ_ALWAYS_INLINE static void PageFree(const Maybe<arena_id_t>& aArenaId,
+                                       void* aPtr) {
+  PtrKind pk = gConst->PtrKind(aPtr);
+  if (pk.IsNothing()) {
+    // Not a page allocation.
+    return aArenaId.isSome() ? sMallocTable.moz_arena_free(*aArenaId, aPtr)
+                             : sMallocTable.free(aPtr);
+  }
+
+  if (pk.IsGuardPage()) {
+    GMut::CrashOnGuardPage(aPtr);
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  // Note that `disable` has no effect unless it is emplaced below.
+  Maybe<AutoDisableOnCurrentThread> disable;
+  // Get the stack trace *before* locking the mutex.
+  StackTrace freeStack;
+  if (GTls::IsDisabledOnCurrentThread()) {
+    // PHC is disabled on this thread. Leave the stack empty.
+  } else {
+    // Disable on this thread *before* getting the stack trace.
+    disable.emplace();
+    freeStack.Fill();
+  }
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  // Check for a double-free.
+  gMut->EnsureValidAndInUse(lock, aPtr, index);
+
+  // Note that FreePage() checks aArenaId (via SetPageFreed()).
+  Delay reuseDelay = ReuseDelay(lock);
+  FreePage(lock, index, aArenaId, freeStack, reuseDelay);
+
+  LOG("PageFree(%p[%zu]), %zu delay, reuse at ~%zu, fullness %zu/%zu\n", aPtr,
+      index, size_t(reuseDelay), size_t(GAtomic::Now()) + reuseDelay,
+      gMut->NumPageAllocs(lock), kNumAllocPages);
+}
+
+static void replace_free(void* aPtr) { return PageFree(Nothing(), aPtr); }
+
+// This handles memalign and moz_arena_memalign.
+MOZ_ALWAYS_INLINE static void* PageMemalign(const Maybe<arena_id_t>& aArenaId,
+                                            size_t aAlignment,
+                                            size_t aReqSize) {
+  MOZ_RELEASE_ASSERT(IsPowerOfTwo(aAlignment));
+
+  // PHC can't satisfy an alignment greater than a page size, so fall back to
+  // mozjemalloc in that case.
+  void* ptr = nullptr;
+  if (aAlignment <= kPageSize) {
+    ptr = MaybePageAlloc(aArenaId, aReqSize, aAlignment, /* aZero */ false);
+  }
+  return ptr ? ptr
+             : (aArenaId.isSome()
+                    ? sMallocTable.moz_arena_memalign(*aArenaId, aAlignment,
+                                                      aReqSize)
+                    : sMallocTable.memalign(aAlignment, aReqSize));
+}
+
+static void* replace_memalign(size_t aAlignment, size_t aReqSize) {
+  return PageMemalign(Nothing(), aAlignment, aReqSize);
+}
+
+static size_t replace_malloc_usable_size(usable_ptr_t aPtr) {
+  PtrKind pk = gConst->PtrKind(aPtr);
+  if (pk.IsNothing()) {
+    // Not a page allocation. Measure it normally.
+    return sMallocTable.malloc_usable_size(aPtr);
+  }
+
+  if (pk.IsGuardPage()) {
+    GMut::CrashOnGuardPage(const_cast<void*>(aPtr));
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  // Check for malloc_usable_size() of a freed block.
+  gMut->EnsureValidAndInUse(lock, const_cast<void*>(aPtr), index);
+
+  return gMut->PageUsableSize(lock, index);
+}
+
+void replace_jemalloc_stats(jemalloc_stats_t* aStats,
+                            jemalloc_bin_stats_t* aBinStats) {
+  sMallocTable.jemalloc_stats_internal(aStats, aBinStats);
+
+  // Add all the pages to `mapped`.
+  size_t mapped = kAllPagesSize;
+  aStats->mapped += mapped;
+
+  size_t allocated = 0;
+  {
+    MutexAutoLock lock(GMut::sMutex);
+
+    // Add usable space of in-use allocations to `allocated`.
+    for (size_t i = 0; i < kNumAllocPages; i++) {
+      if (gMut->IsPageInUse(lock, i)) {
+        allocated += gMut->PageUsableSize(lock, i);
+      }
+    }
+  }
+  aStats->allocated += allocated;
+
+  // Waste is the gap between `allocated` and `mapped`.
+  size_t waste = mapped - allocated;
+  aStats->waste += waste;
+
+  // aStats.page_cache and aStats.bin_unused are left unchanged because PHC
+  // doesn't have anything corresponding to those.
+
+  // gConst and gMut are normal heap allocations, so they're measured by
+  // mozjemalloc as `allocated`. Move them into `bookkeeping`.
+  size_t bookkeeping = sMallocTable.malloc_usable_size(gConst) +
+                       sMallocTable.malloc_usable_size(gMut);
+  aStats->allocated -= bookkeeping;
+  aStats->bookkeeping += bookkeeping;
+}
+
+void replace_jemalloc_ptr_info(const void* aPtr, jemalloc_ptr_info_t* aInfo) {
+  // We need to implement this properly, because various code locations do
+  // things like checking that allocations are in the expected arena.
+  PtrKind pk = gConst->PtrKind(aPtr);
+  if (pk.IsNothing()) {
+    // Not a page allocation.
+    return sMallocTable.jemalloc_ptr_info(aPtr, aInfo);
+  }
+
+  if (pk.IsGuardPage()) {
+    // Treat a guard page as unknown because there's no better alternative.
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+    return;
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  gMut->FillJemallocPtrInfo(lock, aPtr, index, aInfo);
+#if DEBUG
+  LOG("JemallocPtrInfo(%p[%zu]) -> {%zu, %p, %zu, %zu}\n", aPtr, index,
+      size_t(aInfo->tag), aInfo->addr, aInfo->size, aInfo->arenaId);
+#else
+  LOG("JemallocPtrInfo(%p[%zu]) -> {%zu, %p, %zu}\n", aPtr, index,
+      size_t(aInfo->tag), aInfo->addr, aInfo->size);
+#endif
+}
+
+arena_id_t replace_moz_create_arena_with_params(arena_params_t* aParams) {
+  // No need to do anything special here.
+  return sMallocTable.moz_create_arena_with_params(aParams);
+}
+
+void replace_moz_dispose_arena(arena_id_t aArenaId) {
+  // No need to do anything special here.
+  return sMallocTable.moz_dispose_arena(aArenaId);
+}
+
+void* replace_moz_arena_malloc(arena_id_t aArenaId, size_t aReqSize) {
+  return PageMalloc(Some(aArenaId), aReqSize);
+}
+
+void* replace_moz_arena_calloc(arena_id_t aArenaId, size_t aNum,
+                               size_t aReqSize) {
+  return PageCalloc(Some(aArenaId), aNum, aReqSize);
+}
+
+void* replace_moz_arena_realloc(arena_id_t aArenaId, void* aOldPtr,
+                                size_t aNewSize) {
+  return PageRealloc(Some(aArenaId), aOldPtr, aNewSize);
+}
+
+void replace_moz_arena_free(arena_id_t aArenaId, void* aPtr) {
+  return PageFree(Some(aArenaId), aPtr);
+}
+
+void* replace_moz_arena_memalign(arena_id_t aArenaId, size_t aAlignment,
+                                 size_t aReqSize) {
+  return PageMemalign(Some(aArenaId), aAlignment, aReqSize);
+}
+
+class PHCBridge : public ReplaceMallocBridge {
+  virtual bool IsPHCAllocation(const void* aPtr, phc::AddrInfo* aOut) override {
+    PtrKind pk = gConst->PtrKind(aPtr);
+    if (pk.IsNothing()) {
+      return false;
+    }
+
+    bool isGuardPage = false;
+    if (pk.IsGuardPage()) {
+      if ((uintptr_t(aPtr) % kPageSize) < (kPageSize / 2)) {
+        // The address is in the lower half of a guard page, so it's probably an
+        // overflow. But first check that it is not on the very first guard
+        // page, in which case it cannot be an overflow, and we ignore it.
+        if (gConst->IsInFirstGuardPage(aPtr)) {
+          return false;
+        }
+
+        // Get the allocation page preceding this guard page.
+        pk = gConst->PtrKind(static_cast<const uint8_t*>(aPtr) - kPageSize);
+
+      } else {
+        // The address is in the upper half of a guard page, so it's probably an
+        // underflow. Get the allocation page following this guard page.
+        pk = gConst->PtrKind(static_cast<const uint8_t*>(aPtr) + kPageSize);
+      }
+
+      // Make a note of the fact that we hit a guard page.
+      isGuardPage = true;
+    }
+
+    // At this point we know we have an allocation page.
+    uintptr_t index = pk.AllocPageIndex();
+
+    if (aOut) {
+      MutexAutoLock lock(GMut::sMutex);
+      gMut->FillAddrInfo(lock, index, aPtr, isGuardPage, *aOut);
+      LOG("IsPHCAllocation: %zu, %p, %zu, %zu, %zu\n", size_t(aOut->mKind),
+          aOut->mBaseAddr, aOut->mUsableSize,
+          aOut->mAllocStack.isSome() ? aOut->mAllocStack->mLength : 0,
+          aOut->mFreeStack.isSome() ? aOut->mFreeStack->mLength : 0);
+    }
+    return true;
+  }
+
+  virtual void DisablePHCOnCurrentThread() override {
+    GTls::DisableOnCurrentThread();
+    LOG("DisablePHCOnCurrentThread: %zu\n", 0ul);
+  }
+
+  virtual void ReenablePHCOnCurrentThread() override {
+    GTls::EnableOnCurrentThread();
+    LOG("ReenablePHCOnCurrentThread: %zu\n", 0ul);
+  }
+
+  virtual bool IsPHCEnabledOnCurrentThread() override {
+    bool enabled = !GTls::IsDisabledOnCurrentThread();
+    LOG("IsPHCEnabledOnCurrentThread: %zu\n", size_t(enabled));
+    return enabled;
+  }
+};
+
+// WARNING: this function runs *very* early -- before all static initializers
+// have run. For this reason, non-scalar globals (gConst, gMut) are allocated
+// dynamically (so we can guarantee their construction in this function) rather
+// than statically. GAtomic and GTls contain simple static data that doesn't
+// involve static initializers so they don't need to be allocated dynamically.
+void replace_init(malloc_table_t* aMallocTable, ReplaceMallocBridge** aBridge) {
+  // Don't run PHC if the page size isn't 4 KiB.
+  jemalloc_stats_t stats;
+  aMallocTable->jemalloc_stats_internal(&stats, nullptr);
+  if (stats.page_size != kPageSize) {
+    return;
+  }
+
+  sMallocTable = *aMallocTable;
+
+  // The choices of which functions to replace are complex enough that we set
+  // them individually instead of using MALLOC_FUNCS/malloc_decls.h.
+
+  aMallocTable->malloc = replace_malloc;
+  aMallocTable->calloc = replace_calloc;
+  aMallocTable->realloc = replace_realloc;
+  aMallocTable->free = replace_free;
+  aMallocTable->memalign = replace_memalign;
+
+  // posix_memalign, aligned_alloc & valloc: unset, which means they fall back
+  // to replace_memalign.
+  aMallocTable->malloc_usable_size = replace_malloc_usable_size;
+  // default malloc_good_size: the default suffices.
+
+  aMallocTable->jemalloc_stats_internal = replace_jemalloc_stats;
+  // jemalloc_purge_freed_pages: the default suffices.
+  // jemalloc_free_dirty_pages: the default suffices.
+  // jemalloc_thread_local_arena: the default suffices.
+  aMallocTable->jemalloc_ptr_info = replace_jemalloc_ptr_info;
+
+  aMallocTable->moz_create_arena_with_params =
+      replace_moz_create_arena_with_params;
+  aMallocTable->moz_dispose_arena = replace_moz_dispose_arena;
+  aMallocTable->moz_arena_malloc = replace_moz_arena_malloc;
+  aMallocTable->moz_arena_calloc = replace_moz_arena_calloc;
+  aMallocTable->moz_arena_realloc = replace_moz_arena_realloc;
+  aMallocTable->moz_arena_free = replace_moz_arena_free;
+  aMallocTable->moz_arena_memalign = replace_moz_arena_memalign;
+
+  static PHCBridge bridge;
+  *aBridge = &bridge;
+
+#ifndef XP_WIN
+  // Avoid deadlocks when forking by acquiring our state lock prior to forking
+  // and releasing it after forking. See |LogAlloc|'s |replace_init| for
+  // in-depth details.
+  //
+  // Note: This must run after attempting an allocation so as to give the
+  // system malloc a chance to insert its own atfork handler.
+  sMallocTable.malloc(-1);
+  pthread_atfork(GMut::prefork, GMut::postfork, GMut::postfork);
+#endif
+
+  // gConst and gMut are never freed. They live for the life of the process.
+  gConst = InfallibleAllocPolicy::new_<GConst>();
+  GTls::Init();
+  gMut = InfallibleAllocPolicy::new_<GMut>();
+  {
+    MutexAutoLock lock(GMut::sMutex);
+    Delay firstAllocDelay =
+        Rnd64ToDelay<kAvgFirstAllocDelay>(gMut->Random64(lock));
+    GAtomic::Init(firstAllocDelay);
+  }
+}
diff --git a/memory/replace/phc/PHC.h b/memory/replace/phc/PHC.h
new file mode 100644
index 0000000000..139a0e81c9
--- /dev/null
+++ b/memory/replace/phc/PHC.h
@@ -0,0 +1,102 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef PHC_h
+#define PHC_h
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Maybe.h"
+#include <stdint.h>
+#include <stdlib.h>
+
+namespace mozilla {
+namespace phc {
+
+// Note: a stack trace may have no frames due to a collection problem.
+//
+// Also note: a more compact stack trace representation could be achieved with
+// some effort.
+struct StackTrace {
+ public:
+  static const size_t kMaxFrames = 16;
+
+  // The number of PCs in the stack trace.
+  size_t mLength;
+
+  // The PCs in the stack trace. Only the first mLength are initialized.
+  const void* mPcs[kMaxFrames];
+
+ public:
+  StackTrace() : mLength(0) {}
+};
+
+// Info from PHC about an address in memory.
+class AddrInfo {
+ public:
+  enum class Kind {
+    // The address is not in PHC-managed memory.
+    Unknown = 0,
+
+    // The address is within a PHC page that has never been allocated. A crash
+    // involving such an address is unlikely in practice, because it would
+    // require the crash to happen quite early.
+    NeverAllocatedPage = 1,
+
+    // The address is within a PHC page that is in use.
+    InUsePage = 2,
+
+    // The address is within a PHC page that has been allocated and then freed.
+    // A crash involving such an address most likely indicates a
+    // use-after-free. (A sufficiently wild write -- e.g. a large buffer
+    // overflow -- could also trigger it, but this is less likely.)
+    FreedPage = 3,
+
+    // The address is within a PHC guard page. A crash involving such an
+    // address most likely indicates a buffer overflow. (Again, a sufficiently
+    // wild write could unluckily trigger it, but this is less likely.)
+    GuardPage = 4,
+  };
+
+  // The page kind.
+  Kind mKind;
+
+  // The starting address of the allocation.
+  // - Unknown | NeverAllocatedPage: nullptr.
+  // - InUsePage | FreedPage: the address of the allocation within the page.
+  // - GuardPage: the mBaseAddr value from the preceding allocation page.
+  const void* mBaseAddr;
+
+  // The usable size, which could be bigger than the requested size.
+  // - Unknown | NeverAllocatePage: 0.
+  // - InUsePage | FreedPage: the usable size of the allocation within the page.
+  // - GuardPage: the mUsableSize value from the preceding allocation page.
+  size_t mUsableSize;
+
+  // The allocation stack.
+  // - Unknown | NeverAllocatedPage: Nothing.
+  // - InUsePage | FreedPage: Some.
+  // - GuardPage: the mAllocStack value from the preceding allocation page.
+  mozilla::Maybe<StackTrace> mAllocStack;
+
+  // The free stack.
+  // - Unknown | NeverAllocatedPage | InUsePage: Nothing.
+  // - FreedPage: Some.
+  // - GuardPage: the mFreeStack value from the preceding allocation page.
+  mozilla::Maybe<StackTrace> mFreeStack;
+
+  // Default to no PHC info.
+  AddrInfo()
+      : mKind(Kind::Unknown),
+        mBaseAddr(nullptr),
+        mUsableSize(0),
+        mAllocStack(),
+        mFreeStack() {}
+};
+
+}  // namespace phc
+}  // namespace mozilla
+
+#endif /* PHC_h */
diff --git a/memory/replace/phc/moz.build b/memory/replace/phc/moz.build
new file mode 100644
index 0000000000..1a55ab142b
--- /dev/null
+++ b/memory/replace/phc/moz.build
@@ -0,0 +1,36 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+ReplaceMalloc("phc")
+
+DEFINES["MOZ_NO_MOZALLOC"] = True
+
+LOCAL_INCLUDES += [
+    "../logalloc",
+    "/memory/build",
+]
+
+EXPORTS += [
+    "PHC.h",
+]
+
+UNIFIED_SOURCES += [
+    "PHC.cpp",
+]
+
+if not CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    SOURCES += [
+        "../logalloc/FdPrintf.cpp",
+        "/mozglue/misc/StackWalk.cpp",
+    ]
+    if CONFIG["OS_ARCH"] == "WINNT":
+        OS_LIBS += [
+            "dbghelp",
+        ]
+
+TEST_DIRS += ["test"]
+
+DisableStlWrapping()
diff --git a/memory/replace/phc/test/gtest/TestPHC.cpp b/memory/replace/phc/test/gtest/TestPHC.cpp
new file mode 100644
index 0000000000..b83db800f2
--- /dev/null
+++ b/memory/replace/phc/test/gtest/TestPHC.cpp
@@ -0,0 +1,303 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "gtest/gtest.h"
+
+#include "mozmemory.h"
+#include "replace_malloc_bridge.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/mozalloc.h"
+#include "../../PHC.h"
+
+using namespace mozilla;
+
+bool PHCInfoEq(phc::AddrInfo& aInfo, phc::AddrInfo::Kind aKind, void* aBaseAddr,
+               size_t aUsableSize, bool aHasAllocStack, bool aHasFreeStack) {
+  return aInfo.mKind == aKind && aInfo.mBaseAddr == aBaseAddr &&
+         aInfo.mUsableSize == aUsableSize &&
+         // Proper stack traces will have at least 3 elements.
+         (aHasAllocStack ? (aInfo.mAllocStack->mLength > 2)
+                         : (aInfo.mAllocStack.isNothing())) &&
+         (aHasFreeStack ? (aInfo.mFreeStack->mLength > 2)
+                        : (aInfo.mFreeStack.isNothing()));
+}
+
+bool JeInfoEq(jemalloc_ptr_info_t& aInfo, PtrInfoTag aTag, void* aAddr,
+              size_t aSize, arena_id_t arenaId) {
+  return aInfo.tag == aTag && aInfo.addr == aAddr && aInfo.size == aSize
+#ifdef MOZ_DEBUG
+         && aInfo.arenaId == arenaId
+#endif
+      ;
+}
+
+uint8_t* GetPHCAllocation(size_t aSize, size_t aAlignment = 1) {
+  // A crude but effective way to get a PHC allocation.
+  for (int i = 0; i < 2000000; i++) {
+    void* p = (aAlignment == 1) ? moz_xmalloc(aSize)
+                                : moz_xmemalign(aAlignment, aSize);
+    if (ReplaceMalloc::IsPHCAllocation(p, nullptr)) {
+      return (uint8_t*)p;
+    }
+    free(p);
+  }
+  return nullptr;
+}
+
+static const size_t kPageSize = 4096;
+
+TEST(PHC, TestPHCAllocations)
+{
+  // First, check that allocations of various sizes all get put at the end of
+  // their page as expected. Also, check their sizes are as expected.
+
+#define ASSERT_POS(n1, n2)                                      \
+  p = (uint8_t*)moz_xrealloc(p, (n1));                          \
+  ASSERT_EQ((reinterpret_cast<uintptr_t>(p) & (kPageSize - 1)), \
+            kPageSize - (n2));                                  \
+  ASSERT_EQ(moz_malloc_usable_size(p), (n2));
+
+  uint8_t* p = GetPHCAllocation(1);
+  if (!p) {
+    MOZ_CRASH("failed to get a PHC allocation");
+  }
+
+  // On Win64 the smallest possible allocation is 16 bytes. On other platforms
+  // it is 8 bytes.
+#if defined(XP_WIN) && defined(HAVE_64BIT_BUILD)
+  ASSERT_POS(8U, 16U);
+#else
+  ASSERT_POS(8U, 8U);
+#endif
+  ASSERT_POS(16U, 16U);
+  ASSERT_POS(32U, 32U);
+  ASSERT_POS(64U, 64U);
+  ASSERT_POS(128U, 128U);
+  ASSERT_POS(256U, 256U);
+  ASSERT_POS(512U, 512U);
+  ASSERT_POS(1024U, 1024U);
+  ASSERT_POS(2048U, 2048U);
+  ASSERT_POS(4096U, 4096U);
+
+  free(p);
+
+#undef ASSERT_POS
+
+  // Second, do similar checking with allocations of various alignments. Also
+  // check that their sizes (which are different to allocations with normal
+  // alignment) are the same as the sizes of equivalent non-PHC allocations.
+
+#define ASSERT_ALIGN(a1, a2)                                    \
+  p = (uint8_t*)GetPHCAllocation(8, (a1));                      \
+  ASSERT_EQ((reinterpret_cast<uintptr_t>(p) & (kPageSize - 1)), \
+            kPageSize - (a2));                                  \
+  ASSERT_EQ(moz_malloc_usable_size(p), (a2));                   \
+  free(p);                                                      \
+  p = (uint8_t*)moz_xmemalign((a1), 8);                         \
+  ASSERT_EQ(moz_malloc_usable_size(p), (a2));                   \
+  free(p);
+
+  // On Win64 the smallest possible allocation is 16 bytes. On other platforms
+  // it is 8 bytes.
+#if defined(XP_WIN) && defined(HAVE_64BIT_BUILD)
+  ASSERT_ALIGN(8U, 16U);
+#else
+  ASSERT_ALIGN(8U, 8U);
+#endif
+  ASSERT_ALIGN(16U, 16U);
+  ASSERT_ALIGN(32U, 32U);
+  ASSERT_ALIGN(64U, 64U);
+  ASSERT_ALIGN(128U, 128U);
+  ASSERT_ALIGN(256U, 256U);
+  ASSERT_ALIGN(512U, 512U);
+  ASSERT_ALIGN(1024U, 1024U);
+  ASSERT_ALIGN(2048U, 2048U);
+  ASSERT_ALIGN(4096U, 4096U);
+
+#undef ASSERT_ALIGN
+}
+
+TEST(PHC, TestPHCInfo)
+{
+  int stackVar;
+  phc::AddrInfo phcInfo;
+  jemalloc_ptr_info_t jeInfo;
+
+  // Test a default AddrInfo.
+  ASSERT_TRUE(PHCInfoEq(phcInfo, phc::AddrInfo::Kind::Unknown, nullptr, 0ul,
+                        false, false));
+
+  // Test some non-PHC allocation addresses.
+  ASSERT_FALSE(ReplaceMalloc::IsPHCAllocation(nullptr, &phcInfo));
+  ASSERT_TRUE(PHCInfoEq(phcInfo, phc::AddrInfo::Kind::Unknown, nullptr, 0,
+                        false, false));
+  ASSERT_FALSE(ReplaceMalloc::IsPHCAllocation(&stackVar, &phcInfo));
+  ASSERT_TRUE(PHCInfoEq(phcInfo, phc::AddrInfo::Kind::Unknown, nullptr, 0,
+                        false, false));
+
+  uint8_t* p = GetPHCAllocation(32);
+  if (!p) {
+    MOZ_CRASH("failed to get a PHC allocation");
+  }
+
+  // Test an in-use PHC allocation: first byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  ASSERT_EQ(moz_malloc_usable_size(p), 32ul);
+  jemalloc_ptr_info(p, &jeInfo);
+
+  // Test an in-use PHC allocation: last byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagLiveAlloc, p, 32, 0));
+  jemalloc_ptr_info(p + 31, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagLiveAlloc, p, 32, 0));
+
+  // Test an in-use PHC allocation: last byte before it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p - 1, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  jemalloc_ptr_info(p - 1, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: first byte on its allocation page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 32 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: first byte in the following guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 32, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: last byte in the lower half of the
+  // following guard page.
+  ASSERT_TRUE(
+      ReplaceMalloc::IsPHCAllocation(p + 32 + (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 32 + (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: last byte in the preceding guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 31 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: first byte in the upper half of the
+  // preceding guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(
+      p + 31 - kPageSize - (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 31 - kPageSize - (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  free(p);
+
+  // Test a freed PHC allocation: first byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagFreedAlloc, p, 32, 0));
+
+  // Test a freed PHC allocation: last byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 31, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagFreedAlloc, p, 32, 0));
+
+  // Test a freed PHC allocation: last byte before it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p - 1, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p - 1, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: first byte on its allocation page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 32 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: first byte in the following guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 32, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: last byte in the lower half of the following
+  // guard page.
+  ASSERT_TRUE(
+      ReplaceMalloc::IsPHCAllocation(p + 32 + (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 32 + (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: last byte in the preceding guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 31 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: first byte in the upper half of the preceding
+  // guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(
+      p + 31 - kPageSize - (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 31 - kPageSize - (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // There are no tests for `mKind == NeverAllocatedPage` because it's not
+  // possible to reliably get ahold of such a page.
+}
+
+TEST(PHC, TestPHCDisabling)
+{
+  uint8_t* p = GetPHCAllocation(32);
+  uint8_t* q = GetPHCAllocation(32);
+  if (!p || !q) {
+    MOZ_CRASH("failed to get a PHC allocation");
+  }
+
+  ASSERT_TRUE(ReplaceMalloc::IsPHCEnabledOnCurrentThread());
+  ReplaceMalloc::DisablePHCOnCurrentThread();
+  ASSERT_FALSE(ReplaceMalloc::IsPHCEnabledOnCurrentThread());
+
+  // Test realloc() on a PHC allocation while PHC is disabled on the thread.
+  uint8_t* p2 = (uint8_t*)realloc(p, 128);
+  // The small realloc is fulfilled within the same page, but it does move.
+  ASSERT_TRUE(p2 == p - 96);
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p2, nullptr));
+  uint8_t* p3 = (uint8_t*)realloc(p2, 8192);
+  // The big realloc is not in-place, and the result is not a PHC allocation.
+  ASSERT_TRUE(p3 != p2);
+  ASSERT_FALSE(ReplaceMalloc::IsPHCAllocation(p3, nullptr));
+  free(p3);
+
+  // Test free() on a PHC allocation while PHC is disabled on the thread.
+  free(q);
+
+  // These must not be PHC allocations.
+  uint8_t* r = GetPHCAllocation(32);  // This will fail.
+  ASSERT_FALSE(!!r);
+
+  ReplaceMalloc::ReenablePHCOnCurrentThread();
+  ASSERT_TRUE(ReplaceMalloc::IsPHCEnabledOnCurrentThread());
+}
diff --git a/memory/replace/phc/test/gtest/moz.build b/memory/replace/phc/test/gtest/moz.build
new file mode 100644
index 0000000000..82ccaaf9c6
--- /dev/null
+++ b/memory/replace/phc/test/gtest/moz.build
@@ -0,0 +1,15 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+UNIFIED_SOURCES += [
+    "TestPHC.cpp",
+]
+
+LOCAL_INCLUDES += [
+    "../../",
+]
+
+FINAL_LIBRARY = "xul-gtest"
diff --git a/memory/replace/phc/test/moz.build b/memory/replace/phc/test/moz.build
new file mode 100644
index 0000000000..8208ae849d
--- /dev/null
+++ b/memory/replace/phc/test/moz.build
@@ -0,0 +1,9 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+# The gtests won't work in a SpiderMonkey-only build.
+if CONFIG["MOZ_WIDGET_TOOLKIT"]:
+    TEST_DIRS += ["gtest"]